This section reproduces the release notes for new features and incompatible changes in prior releases of Apache Kudu.
|The list of known issues and limitations for prior releases are not reproduced on this page. Please consult the documentation of the appropriate release for a list of known issues and limitations.|
Upgrading directly from Kudu 1.6.0 is supported and no special upgrade steps are required. A rolling upgrade of the server side will not work because the default replica management scheme changed, and running masters and tablet servers with different replica management schemes is not supported, see Incompatible Changes in Kudu 1.7.0 for details. However, mixing client and server sides of different versions is not a problem. You can still update your clients before your servers or vice versa. When upgrading to Kudu 1.7, it is required to first shut down all Kudu processes across the cluster, then upgrade the software on all servers, then restart the Kudu processes on all servers in the cluster.
tcmalloc_contention_time metric, which previously tracked the amount
of time spent in memory allocator lock contention, has been removed.
Support for Java 7 has been deprecated since Kudu 1.5.0 and may be removed in the next major release.
Kudu now supports the decimal column type. The decimal type is a numeric data type with fixed scale and precision suitable for financial and other arithmetic calculations where the imprecise representation and rounding behavior of float and double make those types impractical. The decimal type is also useful for integers larger than int64 and cases with fractional values in a primary key. See Decimal Type for more details.
The strategy Kudu uses for automatically healing tablets which have lost a replica due to server or disk failures has been improved. The new re-replication strategy, or replica management scheme, first adds a replacement tablet replica before evicting the failed one. With the previous replica management scheme, the system first evicts the failed replica and then adds a replacement. The new replica management scheme allows for much faster recovery of tablets in scenarios where one tablet server goes down and then returns back shortly after 5 minutes or so. The new scheme also provides substantially better overall stability on clusters with frequent server failures. (see KUDU-1097).
kudu fs update_dirs tool now supports removing directories. Unless the
--force flag is specified, Kudu will not allow the removal of a directory
across which tablets are configured to spread data. If specified, all tablet
replicas configured to use that directory will fail upon starting up and be
replicated elsewhere, provided a majority exists elsewhere.
Users can use the new
--fs_metadata_dir to specify the directory in which
to place tablet-specific metadata. It is recommended, although not
necessary, that this be placed on a high-performance drive with high
bandwidth and low latency, e.g. a solid-state drive. If not specified,
metadata will be placed in the directory specified by
--fs_wal_dir, or in
the directory specified by the first entry of
--fs_data_dirs if metadata
already exists there from a pre-Kudu 1.7 deployment. Kudu will not
automatically move existing metadata based on this configuration.
Kudu 1.7 introduces a new scan read mode READ_YOUR_WRITES. Users can specify READ_YOUR_WRITES when creating a new scanner in C++, Java and Python clients. If this mode is used, the client will perform a read such that it follows all previously known writes and reads from this client. Reads in this mode ensure read-your-writes and read-your-reads session guarantees, while minimizing latency caused by waiting for outstanding write transactions to complete. Note that this is still an experimental feature which may be stabilized in future releases.
The tablet server web UI scans dashboard (/scans) has been improved with several new features, including: showing the most recently completed scans, a pseudo-SQL scan descriptor that concisely shows the selected columns and applied predicates, and more complete and better documented scan statistics.
Kudu daemons now expose a web page
/stacks which dumps the current stack
trace of every thread running in the server. This information can be helpful
when diagnosing performance issues.
By default, each tablet replica will now stripe data blocks across 3 data directories instead of all data directories. This decreases the likelihood that any given tablet will be affected in the event of a single disk failure. No substantial performance impact is expected due to this feature based on performance testing. This change only affects new replicas created after upgrading to Kudu 1.7.
Kudu servers previously offered the ability to enable a separate metrics log which stores periodic snapshots of all metrics available on a server. This functionality is now available as part of a more general “diagnostics log” which is enabled by default. The diagnostics log includes periodic dumps of server metrics as well as collections of thread stack traces. The default configuration ensures that no more than 640MB of diagnostics logs are retained, and typically the space consumption is significantly less due to compression. The format and contents of this log file are documented in the Administration guide.
The handling of errors in the synchronous Java client has been improved so that, when an exception is thrown, the stack trace indicates the correct location where the client function was invoked rather than a call stack of an internal worker thread. The original call stack from the worker thread is available as a “suppressed exception”.
The logging of errors in the Java client has been improved to exclude exception stack traces for expected scenarios such as failure to connect to a server in a cluster. Instead, only a single line informational message will be logged in such cases to aid in debugging.
The Java client now uses a predefined prioritized list of TLS ciphers when establishing an encrypted connection to Kudu servers. This cipher list matches the list of ciphers preferred for server-to-server communication and ensures that the most efficient and secure ciphers are preferred. When the Kudu client is running on Java 8 or newer, this provides a substantial speed-up to read and write performance.
Reporting for the
kudu cluster ksck tool has been updated so tablets and
tables with on-going tablet copies are shown as "recovering". Additional
reporting changes have been made to make various common scenarios,
particularly tablet copies, less alarming.
The performance of inserting rows containing many string or binary columns has been improved, especially in the case of highly concurrent write workloads.
By default, Spark tasks that scan Kudu will now be able to scan non-leader replicas. This allows Spark to more easily schedule kudu-spark tasks local to the data. Users can disable this behavior by passing 'leader_only' to the 'kudu.scanLocality' option."
The number of OS threads used in the steady state and during bursts of
activity (such as in Raft leader elections triggered by a node failure) has
been drastically reduced and should no longer exceed the value of
As such, it should no longer be necessary to increase the value of
(or of /proc/sys/kernel/threads-max) in order to run a Kudu tablet server in
An issue where sparse column predicates could cause excessive data-block reads has been fixed. Previously in certain scans with sparsely matching predicates on multiple columns, Kudu would read and decode the same data blocks many times. The improvement typically results in a 5-10x performance increase for the affected scans. (see KUDU-2231).
The efficiency and on-disk size of large updated values has been improved. This will improve update-heavy workloads which overwrite large (1KiB+) values. (see KUDU-2253).
Fixed a scenario where the on-disk data of a tablet server was completely erased and and a new tablet server was started on the same host. This issue could prevent tablet replicas previously hosted on the server from being evicted and re-replicated. Tablets now immediately evict replicas that respond with a different server UUID than expected. (see KUDU-1613).
Fixed a rare race condition when connecting to masters during their startup which might cause a client to get a response without a CA certificate and/or authentication token. This would cause the client to fail to authenticate with other servers in the cluster. The leader master now always sends a CA certificate and an authentication token (when applicable) to a Kudu client with a successful ConnectToMaster response. (see KUDU-1927).
The Kudu Java client now will retry a connection if no master is discovered as a leader, and the user has a valid authentication token. This avoids failure in recoverable cases when masters are in the process of the very first leader election after starting up. (see KUDU-2262).
The Java client will now automatically attempt to re-acquire Kerberos
credentials from the ticket cache when the prior credentials are about to
expire. This allows client instances to persist longer than the expiration
time of a single Kerberos ticket so long as some other process renews the
credentials in the ticket cache. Documentation on interacting with Kerberos
authentication has been added to the Javadoc for the
Follower masters are now able to verify authentication tokens even if they have never been a leader. Prior to this fix, if a follower master had never been a leader, clients would be unable to authenticate to that master, resulting in spurious error messages being logged. (see KUDU-2265).
Fixed a tablet server crash when a tablet replica is deleted during a scan. (see KUDU-2295).
The evaluation order of predicates in scans with multiple predicates has been made deterministic. Due to a bug, this was not necessarily the case previously. Predicates are applied in most to least selective order, with ties broken by column index. The evaluation order may change in the future, particularly when better column statistics are made available internally. (see KUDU-2312).
kudu tablet change_config move_replica tool required all
tablet servers in the cluster to be available when performing a move. This
restriction has been relaxed: only the tablet server that will receive a replica
of the tablet being moved and the hosts of the tablet’s existing replicas need to be
available for the move to occur.
Fixed a bug in the Java client which prevented the client from locating the new leader master after a leader failover in the case that the previous leader either remained online or restarted quickly. This bug resulted in the client timing out operations with errors indicating that there was no leader master. (see KUDU-2343).
The Unix process username of the client is now included inside the exported
security credentials, so that the effective username of clients who import
credentials and subsequently use unauthenticated (SASL PLAIN) connections
matches the client who exported the security credentials. For example, this is
useful to let the Spark executors know which username to use if the Spark
driver has no authentication token. This change only affects clusters with
encryption disabled using
Kudu 1.7.0 is wire-compatible with previous versions of Kudu:
Kudu 1.7 clients may connect to servers running Kudu 1.0 or later. If the client uses features that are not available on the target server, an error will be returned.
Rolling upgrade between Kudu 1.6 and Kudu 1.7 servers is believed to be possible though has not been sufficiently tested. Users are encouraged to shut down all nodes in the cluster, upgrade the software, and then restart the daemons on the new version.
Kudu 1.0 clients may connect to servers running Kudu 1.7 with the exception of the below-mentioned restrictions regarding secure clusters.
The authentication features introduced in Kudu 1.3 place the following limitations on wire compatibility between Kudu 1.7 and versions earlier than 1.3:
If a Kudu 1.7 cluster is configured with authentication or encryption set to "required", clients older than Kudu 1.3 will be unable to connect.
If a Kudu 1.7 cluster is configured with authentication and encryption set to "optional" or "disabled", older clients will still be able to connect.
The newly introduced replica management scheme is not compatible with the old scheme, so it’s not possible to run pre-1.7 Kudu masters with 1.7 Kudu tablet servers or vice versa. This is a server-side incompatibility only and it does not affect client compatibility. In other words, Kudu clients of prior versions are compatible with upgraded Kudu clusters.
Kudu masters of 1.7 version will not register Kudu tablet servers of 1.6 and prior versions.
Kudu tablet servers of 1.7 version will not work with Kudu masters of 1.6 and prior versions.
The format of the previously-optional metrics log has changed to include a human-readable timestamp on each line. The path of the log file has also changed with the word “diagnostics” replacing the word “metrics” in the file name. The metrics log has been optimized to only include those metrics which have changed in between successive samples, and to not include entity attributes such as tablet partition information in the log. (see KUDU-2297).
The Kudu 1.7 Java client library is API- and ABI-compatible with Kudu 1.6. Applications written against Kudu 1.6 will compile and run against the Kudu 1.7 client library and vice-versa.
The Kudu 1.7 C++ client is API- and ABI-forward-compatible with Kudu 1.6. Applications written and compiled against the Kudu 1.6 client library will run without modification against the Kudu 1.7 client library. Applications written and compiled against the Kudu 1.7 client library will run without modification against the Kudu 1.6 client library.
The Kudu 1.7 Python client is API-compatible with Kudu 1.6. Applications written against Kudu 1.6 will continue to run against the Kudu 1.7 client and vice-versa.
Kudu 1.7 clients that attempt to create a table with a decimal column on a target server running Kudu 1.6 or earlier will receive an error response. Similarly Kudu clients running Kudu 1.6 or earlier will result in an error when attempting to access any table containing containing a decimal column.
Upgrading directly from Kudu 1.5.0 is supported and no special upgrade steps are required. A rolling upgrade may work, however it has not been tested. When upgrading Kudu, it is recommended to first shut down all Kudu processes across the cluster, then upgrade the software on all servers, then restart the Kudu processes on all servers in the cluster.
Support for Spark 1 (kudu-spark_2.10) has been removed in Kudu 1.6.0 and now only Spark 2 is supported. Spark 1 support was deprecated in Kudu 1.5.0.
Support for Java 7 has been deprecated since Kudu 1.5.0 and may be removed in the next major release.
Tablet servers' tolerance of disk failures is now enabled by default and has
been extended to handle data directory failures at runtime. In the event of
a disk failure at runtime, any tablets with data on a failed disk will be
shut down and restarted on another tablet server. There is a configurable
tradeoff between a newly added tablet’s tolerance to disk failures and its
ability to parallelize reads via the experimental
--fs_target_data_dirs_per_tablet flag. Tablets that are spread across fewer
disks are less likely to be affected by a disk failure, at the cost of
reduced parallelism. By default, tablets are striped across all available
disks. Note that the first configured data directory and the WAL directory
cannot currently tolerate disk failures. This will be further improved in
future Kudu releases.
Kudu servers can now adopt new data directories via the new
kudu fs update_dirs tool. The new directory will be used by new tablet
replicas only. Note that removing directories is not yet supported
Kudu servers have two new flags to control webui TLS/HTTPS
These flags allow the advertised TLS ciphers and TLS protocol versions to be
configured. Additionally, the webserver now excludes insecure legacy ciphers
Kudu servers can now tolerate short interruptions in NTP clock synchronization. NTP synchronization is still required when any Kudu daemon starts up. If NTP synchronization is not available, diagnostic information is now logged to help pinpoint the issue (see KUDU-1578).
Tablet server startup time has been improved significantly on servers containing large numbers of blocks.
The log block manager now performs disk data deletion in batches. This optimization can significantly reduce the time taken to delete data on a tablet.
The usage of sensitive data redaction flag has been slightly changed. By
--redact=log flag, redaction will be disabled in the web UI but
retained for server logs. Alternatively,
--redact=none can be used to
disable redaction completely.
The Spark DataSource integration now can take advantage of scan locality for better scan performance, the scan will take place at the closest replica instead of going to the leader.
Various optimizations were made to reduce the 99th percentile latency of writes on the tablet server. This can also improve throughput on certain write workloads, particularly on larger clusters.
Kudu may now be configured to ignore system-wide auth_to_local mappings
configured in /etc/krb5.conf by setting the configuration flag
The performance of the compaction scheduler has been improved. In previous versions, certain types of time series workloads were found to cause compaction scheduling to take tens of seconds. These workloads now schedule compactions an order of magnitude more efficiently.
The compaction scheduler has been improved to avoid running a compaction when the benefit of that compaction is extremely small.
Tablet servers now consider the health of all replicas of a tablet before deciding to evict one. This can improve stability of the Kudu cluster after experiencing multiple simultaneous daemon failures (see KUDU-2048).
Several performance improvements have been made to the Kudu master, particularly in concurrency of clients opening tables. This should improve performance in highly concurrent workloads.
The on-disk size metric for a tablet now includes all data and metadata. Previously, it excluded WAL segments and consensus metadata (see KUDU-1755).
Added verbose mode for the 'kudu cluster ksck' command to enable output of detailed information on the cluster’s metadata, even when no errors are detected.
HybridTime timestamp propagation now works in the Java client when using scan tokens (see KUDU-1411).
Fixed an error message commonly found in tablet server logs indicating that operations were being read "from the future" (see KUDU-1078).
Tombstoned tablets no longer report metrics (see KUDU-2044).
Fixed a bug in the C++ client which could cause tablets to be erroneously pruned, or skipped, during certain scans, resulting in fewer results than expected being returned from queries. The bug only affected tables whose range partition columns are a proper prefix of the primary key (see KUDU-2173).
Published Kudu Java artifacts are now fully compatible with JRE 7 and JRE 8. There was previously a bug in the release process which made them compatible only with JRE 8 (see KUDU-2188).
Fixed a typo in the list of default TLS ciphers used by Kudu servers. As a result, two additional cipher suites are now available:
ECDHE-RSA-AES128-SHA256 TLSv1.2 Kx=ECDH Au=RSA Enc=AES(128) Mac=SHA256
AES256-GCM-SHA384 TLSv1.2 Kx=RSA Au=RSA Enc=AESGCM(256) Mac=AEAD
Kudu 1.6.0 is wire-compatible with previous versions of Kudu:
Kudu 1.6 clients may connect to servers running Kudu 1.0 or later. If the client uses features that are not available on the target server, an error will be returned.
Rolling upgrade between Kudu 1.5 and Kudu 1.6 servers is believed to be possible though has not been sufficiently tested. Users are encouraged to shut down all nodes in the cluster, upgrade the software, and then restart the daemons on the new version.
Kudu 1.0 clients may connect to servers running Kudu 1.6 with the exception of the below-mentioned restrictions regarding secure clusters.
The authentication features introduced in Kudu 1.3 place the following limitations on wire compatibility between Kudu 1.6 and versions earlier than 1.3:
If a Kudu 1.6 cluster is configured with authentication or encryption set to "required", clients older than Kudu 1.3 will be unable to connect.
If a Kudu 1.6 cluster is configured with authentication and encryption set to "optional" or "disabled", older clients will still be able to connect.
The Kudu 1.6 Java client library is API- and ABI-compatible with Kudu 1.5. Applications written against Kudu 1.5 will compile and run against the Kudu 1.6 client library and vice-versa.
The Kudu 1.6 C++ client is API- and ABI-forward-compatible with Kudu 1.5. Applications written and compiled against the Kudu 1.5 client library will run without modification against the Kudu 1.6 client library. Applications written and compiled against the Kudu 1.6 client library will run without modification against the Kudu 1.5 client library.
The Kudu 1.6 Python client is API-compatible with Kudu 1.5. Applications written against Kudu 1.5 will continue to run against the Kudu 1.6 client and vice-versa.
Kudu 1.5 now enables the optional ability to compute, store, and verify checksums on all pieces of data stored on a server by default. Due to storage format changes, downgrading to versions 1.3 or earlier is not supported and will result in an error.
Spark 2.2+ requires Java 8 at runtime even though Kudu Spark 2.x integration is Java 7 compatible. Spark 2.2 is the default dependency version as of Kudu 1.5.0.
The kudu-spark-tools module has been renamed to kudu-spark2-tools_2.11 in order to include the Spark and Scala base versions. This matches the pattern used in the kudu-spark module and artifacts.
To improve security, world-readable Kerberos keytab files are no longer
accepted by default. Set
--allow_world_readable_credentials=true to override
this behavior. See
KUDU-1955 for additional
Support for Java 7 is deprecated as of Kudu 1.5.0 and may be removed in the next major release.
Support for Spark 1 (kudu-spark_2.10) is deprecated as of Kudu 1.5.0 and may be removed in the next minor release.
Tablet servers are now optionally able to tolerate disk failures at
startup. This feature is experimental; by default, Kudu will crash if it
experiences a disk failure. When enabled, tablets with any data on the failed
disk will not be opened and will be replicated as needed. To enable this, set
--crash_on_eio flag to
false. Additionally, there is a configurable
tradeoff between a newly added tablet’s tolerance to disk failures and its
parallelization of I/O via the
Tablets that are spread across fewer disks are less likely to be affected by a
disk failure, at the cost of reduced parallelism. Note that the first
configured data directory and the WAL directory cannot currently tolerate disk
failures, and disk failures during run-time are still fatal.
Kudu server web UIs have a new configuration dashboard (/config) which provides a high level summary of important security configuration values, such as whether RPC authentication is required, or web server HTTPS encryption is enabled. Other types of configuration will be added in future releases.
kudu command line tool has two new features:
kudu tablet change_config
kudu local_replica data_size. The 'tablet change_config
move_replica' tool moves a tablet replica from one tablet server to another,
under the condition that the tablet is healthy. An operator can use this tool to
rebalance tablet replicas between tablet servers. The 'local_replica data size'
tool summarizes the space usage of a tablet, breaking it down by type of file,
column, and rowset.
kudu-client-tools now supports exporting CSV files and importing Apache Parquet files. This feature is unstable and may change APIs and functionality in future releases.
kudu-spark-tools now supports importing and exporting CSV, Apache Avro and Apache Parquet files. This feature is unstable and may change APIs and functionality in future releases.
The log block manager now performs disk synchronization in batches. This optimization can significantly reduce the time taken to copy tablet data from one server to another; in one case tablet copy time is reduced by 35%. It also improves the general performance of flushes and compactions.
A new feature referred to as "tombstoned voting" is added to the Raft
consensus subsystem to allow tablet replicas in the
state to vote in tablet leader elections. This feature increases Kudu’s
stability and availability by improving the likelihood that Kudu will be able
to self-heal in more edge-case scenarios, such as when tablet copy operations
fail. See KUDU-871 for
The tablet on-disk size metric has been made more accurate. Previously, the metric included only REDO deltas; it now counts all deltas. Additionally, the metric includes the size of bloomfiles, ad hoc indexes, and the tablet superblock. WAL segments and consensus metadata are still not counted. The latter is very small compared to the size of data, but the former may be significant depending on the workload (this will be resolved in a future release).
The number of threads used by the Kudu tablet server has been further reduced. Previously, each follower tablet replica used a dedicated thread to detect leader tablet replica failures, and each leader replica used one dedicated thread per follower to send Raft heartbeats to that follower. The work performed by these dedicated threads has been reassigned to other threads. Other improvements were made to facilitate better thread sharing by tablets. For the purpose of capacity planning, expect the Kudu tablet server to create one thread for every five "cold" (i.e. those not servicing writes) tablets, and an additional three threads for every "hot" tablet. This will be further improved upon in future Kudu releases.
The Java Kudu client now automatically requests new authentication tokens after expiration. As a result, long-lived Java clients are now supported. See KUDU-2013 for more details.
Multiple Kerberos compatibility bugs have been fixed, including support for environments with disabled reverse DNS, FreeIPA compatibility, principal names including uppercase characters, and hosts without a FQDN.
A bug in the binary prefix decoder which could cause a tablet server 'check' assertion crash has been fixed. The crash could only be triggered in very specific scenarios; see KUDU-2085 for additional details.
Kudu 1.5.0 is wire-compatible with previous versions of Kudu:
Kudu 1.5 clients may connect to servers running Kudu 1.0 or later. If the client uses features that are not available on the target server, an error will be returned.
Rolling upgrade between Kudu 1.4 and Kudu 1.5 servers is believed to be possible though has not been sufficiently tested. Users are encouraged to shut down all nodes in the cluster, upgrade the software, and then restart the daemons on the new version.
Kudu 1.0 clients may connect to servers running Kudu 1.5 with the exception of the below-mentioned restrictions regarding secure clusters.
The authentication features introduced in Kudu 1.3 place the following limitations on wire compatibility between Kudu 1.5 and versions earlier than 1.3:
If a Kudu 1.5 cluster is configured with authentication or encryption set to "required", clients older than Kudu 1.3 will be unable to connect.
If a Kudu 1.5 cluster is configured with authentication and encryption set to "optional" or "disabled", older clients will still be able to connect.
The Kudu 1.5 Java client library is API- and ABI-compatible with Kudu 1.4. Applications written against Kudu 1.4 will compile and run against the Kudu 1.5 client library and vice-versa, unless one of the following newly added APIs is used:
The Kudu 1.5 C++ client is API- and ABI-forward-compatible with Kudu 1.4. Applications written and compiled against the Kudu 1.4 client library will run without modification against the Kudu 1.5 client library. Applications written and compiled against the Kudu 1.5 client library will run without modification against the Kudu 1.4 client library.
The Kudu 1.5 Python client is API-compatible with Kudu 1.4. Applications written against Kudu 1.4 will continue to run against the Kudu 1.5 client and vice-versa.
The Maintenance Manager now fully uses the threads it’s given (see the improvements described further below), so it’s now able to generate a lot more IO by flushing and compacting more often. Generally, the recommended ratio of MM threads to data directories is 1:3; operators of clusters above that ratio should be mindful of this when upgrading.
The C++ and Java client libraries now support the ability to alter the storage attributes (e.g. encoding and compression) and default value of existing columns. Additionally, it is now possible to rename a column which is part of a table’s primary key.
The C++ client library now includes an experimental
KuduPartitioner API which may
be used to efficiently map rows to their associated partitions and hosts.
This may be used to achieve better locality or distribution of writes
in client applications.
The Java client library now supports enabling fault tolerance on scanners. Fault tolerant scanners are able to transparently recover from concurrent server crashes at the cost of some performance overhead. See the Java API documentation for more details on usage.
kudu command line tool now includes a new advanced administrative
kudu remote_replica unsafe_change_config. This command may be used
to force a tablet to perform an unsafe change of its Raft replication
configuration. This can be used to recover from scenarios such as a loss
of a majority of replicas, at the risk of losing edits.
kudu command line tool now includes the
kudu fs check command
which performs various offline consistency checks on the local on-disk
storage of a Kudu Tablet Server or Master. In addition to detecting
various inconsistencies or corruptions, it can also detect and remove
data blocks that are no longer referenced by any tablet but were not
fully removed from disk due to a crash or a bug in prior versions of Kudu.
kudu command line tool can now be used to list the addresses and
identifiers of the servers in the cluster using either
kudu master list
kudu tserver list.
Kudu 1.4 now includes the optional ability to compute, store, and verify checksums on all pieces of data stored on a server. Prior versions only performed checksums on certain portions of the stored data. This feature is not enabled by default since it makes a backward-incompatible change to the on-disk formats and thus prevent downgrades. Kudu 1.5 will enable the feature by default.
kudu cluster ksck now detects and reports new classes of
inconsistencies and issues. In particular, it is better able to
detect cases where a configuration change such as a replica eviction
or addition is pending but is unable to be committed. It also now
properly detects and reports cases where a tablet has no elected
The default size for Write Ahead Log (WAL) segments has been reduced from 64MB to 8MB. Additionally, in the case that all replicas of a tablet are fully up to date and data has been flushed from memory, servers will now retain only a single WAL segment rather than two. These changes are expected to reduce the average consumption of disk space on the configured WAL disk by 16x, as well as improve the startup speed of tablet servers by reducing the number and size of WAL segments that need to be re-read.
The default on-disk storage system used by Kudu servers (Log Block Manager) has been improved to compact its metadata and remove dead containers. This compaction and garbage collection occurs only at startup. Thus, the first startup after upgrade is expected to be longer than usual, and subsequent restarts should be shorter.
The usability of the Kudu web interfaces has been improved, particularly for the case where a server hosts many tablets or a table has many partitions. Pages that list tablets now include a top-level summary of tablet status and show the complete list under a toggleable section.
The Maintenance Manager has been improved to improve utilization of the configured maintenance threads. Previously, maintenance work would only be scheduled a maximum of 4 times per second, but now maintenance work will be scheduled immediately whenever any configured thread is available. This can improve the throughput of write-heavy workloads.
The Maintenance Manager will now aggressively schedule flushes of in-memory data when memory consumption crosses 60% of the configured process-wide memory limit. The backpressure mechanism which begins to throttle client writes has been accordingly adjusted to not begin throttling until reaching 80% of the configured limit. These two changes together result in improved write throughput, more consistent latency, and fewer timeouts due to memory exhaustion.
Many performance improvements were made to write performance. Applications which send large batches of writes to Kudu should see substantially improved throughput in Kudu 1.4.
Several improvements were made to reduce the memory consumption of Kudu Tablet Servers which hold large volumes of data. The specific amount of memory saved varies depending on workload, but the expectation is that approximately 350MB of excess peak memory usage has been eliminated per TB of data stored.
The number of threads used by the Kudu Tablet Server has been reduced. Previously, each tablet used a dedicated thread to append to its WAL. Those threads now automatically stop running if there is no activity on a given tablet for a short period of time.
KUDU-2020 Fixed an issue where re-replication after a failure would proceed significantly slower than expected. This bug caused many tablets to be unnecessarily copied multiple times before successfully being considered re-replicated, resulting in significantly more network and IO bandwidth usage than expected. Mean time to recovery on clusters with large amounts of data is improved by up to 10x by this fix.
Fixed an issue where the Java client would call
very often, causing performance problems particularly on Windows
where this function can be quite slow.
Improved the accuracy of the
on_disk_size replica metrics to
include the size consumed by bloom filters, primary key indexes,
and superblock metadata, and delta files. Note that, because the size
metric is now more accurate, the reported values are expected to
increase after upgrading to Kudu 1.4. This does not indicate that
replicas are using more space after the upgrade; rather, it is
now accurately reporting the amount of space that has always been
Kudu servers will now periodically flush their log messages to disk
even if no
WARNING-level messages have been logged. This makes it
easier to tail the logs to see progress output during normal startup.
KUDU-1999 Fixed the ability to run Spark jobs in "cluster" mode against Kudu clusters secured by Kerberos.
Kudu 1.4.0 is wire-compatible with previous versions of Kudu:
Kudu 1.4 clients may connect to servers running Kudu 1.0 or later. If the client uses features that are not available on the target server, an error will be returned.
Kudu 1.0 clients may connect to servers running Kudu 1.4 with the exception of the below-mentioned restrictions regarding secure clusters.
Rolling upgrade between Kudu 1.3 and Kudu 1.4 servers is believed to be possible though has not been sufficiently tested. Users are encouraged to shut down all nodes in the cluster, upgrade the software, and then restart the daemons on the new version.
The authentication features introduced in Kudu 1.3 place the following limitations on wire compatibility between Kudu 1.4 and versions earlier than 1.3:
If a Kudu 1.4 cluster is configured with authentication or encryption set to "required", clients older than Kudu 1.3 will be unable to connect.
If a Kudu 1.4 cluster is configured with authentication and encryption set to "optional" or "disabled", older clients will still be able to connect.
Kudu servers, by default, will now only allow unencrypted or unauthenticated connections from trusted subnets, which are private networks (127.0.0.0/8,10.0.0.0/8,172.16.0.0/12, 192.168.0.0/16,169.254.0.0/16) and local subnets of all local network interfaces. Unencrypted or unauthenticated connections from publicly routable IPs will be rejected, even if encryption and authentication are not configured.
The trusted subnets can be configured using the
--trusted_subnets flag, which can be set
to IP blocks represented in CIDR notation separated by comma. Set it to '0.0.0.0/0' to
allow unauthenticated connections from all remote IP addresses. However, if network access
is not otherwise restricted by a firewall, malicious users may be able to gain unauthorized
access. This can be mitigated if authentication and encryption are configured to be
The Kudu 1.4 Java client library is API- and ABI-compatible with Kudu 1.3. Applications written against Kudu 1.3 will compile and run against the Kudu 1.4 client library and vice-versa, unless one of the following newly added APIs is used:
New methods in
New getters in
The Kudu 1.4 C++ client is API- and ABI-forward-compatible with Kudu 1.3. Applications written and compiled against the Kudu 1.3 client library will run without modification against the Kudu 1.4 client library. Applications written and compiled against the Kudu 1.4 client library will run without modification against the Kudu 1.3 client library unless they use one of the following new APIs:
KuduScanner::SetRowFormatFlags (unstable API)
KuduScanBatch::indirect_data (unstable API)
The Kudu 1.4 Python client is API-compatible with Kudu 1.3. Applications written against Kudu 1.3 will continue to run against the Kudu 1.4 client and vice-versa.
Kudu 1.3 adds support for strong authentication based on Kerberos. This optional feature allows users to authenticate themselves using Kerberos tickets, and also provides mutual authentication of servers using Kerberos credentials stored in keytabs. This feature is optional, but recommended for deployments requiring security.
Kudu 1.3 adds support for encryption of data on the network using Transport Layer Security (TLS). Kudu will now use TLS to encrypt all network traffic between clients and servers as well as any internal traffic among servers, with the exception of traffic determined to be within a localhost network connection. Encryption is enabled by default whenever it can be determined that both the client and server support the feature.
Kudu 1.3 adds coarse-grained service-level authorization of access to the cluster. The operator may set up lists of permitted users who may act as administrators and as clients of the cluster. Combined with the strong authentication feature described above, this can enable a secure environment for some use cases. Note that fine-grained access control (e.g. table-level or column-level) is not yet supported.
Kudu 1.3 adds a background task to tablet servers which removes historical versions of data which have fallen behind the configured data retention time. This reduces disk space usage in all workloads, but particularly in those with a higher volume of updates or upserts.
Kudu now incorporates Google Breakpad, a library which writes crash reports in the case of a server crash. These reports can be found within the configured log directory, and can be useful during bug diagnosis.
Kudu servers will now change the file permissions of data directories and contained
data files based on a new configuration flag
--umask. As a result, after upgrading,
permissions on disk may be more restrictive than in previous versions. The new default
configuration improves data security.
Kudu’s web UI will now redact strings which may include sensitive user data. For example, the monitoring page which shows in-progress scans no longer includes the scanner predicate values. The tracing and RPC diagnostics endpoints no longer include contents of RPCs which may include table data.
By default, Kudu now reserves 1% of each configured data volume as free space. If a volume is seen to have less than 1% of disk space free, Kudu will stop writing to that volume to avoid completely filling up the disk.
The default encoding for numeric columns (int, float, and double) has been changed
BIT_SHUFFLE. The default encoding for binary and string columns has been
DICT_ENCODING. Dictionary encoding automatically falls back to the old
PLAIN) when cardinality is too high to be effectively encoded.
These new defaults match the default behavior of other storage mechanisms such as Apache Parquet and are likely to perform better out of the box.
Kudu now uses
LZ4 compression when writing its Write Ahead Log (WAL). This improves
write performance and stability for many use cases.
Kudu now uses
LZ4 compression when writing delta files. This can improve both
read and write performance as well as save substantial disk usage, especially
for workloads involving a high number of updates or upserts containing compressible
The Kudu API now supports the ability to express
IS NULL and
IS NOT NULL predicates
on scanners. The Spark DataSource integration will take advantage of these new
predicates when possible.
Both C++ and Java clients have been optimized to prune partitions more effectively
when performing scans using the
IN (…) predicate.
The exception messages produced by the Java client are now truncated to a maximum length of 32KB.
Fixed a critical bug in which wrong results would be returned when evaluating
predicates applied to columns added using the
ALTER TABLE operation.
KUDU-1905 Fixed a crash after inserting a row sharing a primary key with a recently-deleted row in tables where the primary key is comprised of all of the columns.
KUDU-1899 Fixed a crash after inserting a row with an empty string as the single-column primary key.
KUDU-1904 Fixed a potential crash when performing random reads against a column using RLE encoding and containing long runs of NULL values.
KUDU-1853 Fixed an issue where disk space could be leaked on servers which experienced an error during the process of copying tablet data from another server.
KUDU-1856 Fixed an issue in which disk space could be leaked by Kudu servers storing data on partitions using the XFS file system. Any leaked disk space will be automatically recovered upon upgrade.
Kudu 1.3.0 is wire-compatible with previous versions of Kudu:
Kudu 1.3 clients may connect to servers running Kudu 1.0. If the client uses features that are not available on the target server, an error will be returned.
Kudu 1.0 clients may connect to servers running Kudu 1.3 with the exception of the below-mentioned restrictions regarding secure clusters.
Rolling upgrade between Kudu 1.2 and Kudu 1.3 servers is believed to be possible though has not been sufficiently tested. Users are encouraged to shut down all nodes in the cluster, upgrade the software, and then restart the daemons on the new version.
The authentication features newly introduced in Kudu 1.3 place the following limitations on wire compatibility with older versions:
If a Kudu 1.3 cluster is configured with authentication or encryption set to "required", older clients will be unable to connect.
If a Kudu 1.3 cluster is configured with authentication and encryption set to "optional" or "disabled", older clients will still be able to connect.
Due to storage format changes in Kudu 1.3, downgrade from Kudu 1.3 to earlier versions is not supported. After upgrading to Kudu 1.3, attempting to restart with an earlier version will result in an error.
In order to support running MapReduce and Spark jobs on secure clusters, these frameworks now connect to the cluster at job submission time to retrieve authentication credentials which can later be used by the tasks to be spawned. This means that the process submitting jobs to Kudu clusters must have direct access to that cluster.
The embedded web servers in Kudu processes now specify the
X-Frame-Options: DENY HTTP
header which prevents embedding Kudu web pages in HTML
The Kudu 1.3 Java client library is API- and ABI-compatible with Kudu 1.2. Applications written against Kudu 1.2 will compile and run against the Kudu 1.3 client library and vice-versa, unless one of the following newly added APIs is used:
[Async]KuduClient.exportAuthenticationCredentials(…) (unstable API)
[Async]KuduClient.importAuthenticationCredentials(…) (unstable API)
The Kudu 1.3 C++ client is API- and ABI-forward-compatible with Kudu 1.2. Applications written and compiled against the Kudu 1.2 client library will run without modification against the Kudu 1.3 client library. Applications written and compiled against the Kudu 1.3 client library will run without modification against the Kudu 1.2 client library unless they use one of the following new APIs:
The Kudu 1.3 Python client is API-compatible with Kudu 1.2. Applications written against Kudu 1.2 will continue to run against the Kudu 1.3 client and vice-versa.
Kudu clients and servers now redact user data such as cell values
from log messages, Java exception messages, and
User metadata such as table names, column names, and partition
bounds are not redacted.
Redaction is enabled by default, but may be disabled by setting the new
log_redact_user_data flag to
Kudu’s ability to provide consistency guarantees has been substantially improved:
Replicas now correctly track their "safe timestamp". This timestamp is the maximum timestamp at which reads are guaranteed to be repeatable.
A scan created using the
SCAN_AT_SNAPSHOT mode will now
either wait for the requested snapshot to be "safe" at the replica
being scanned, or be re-routed to a replica where the requested
snapshot is "safe". This ensures that all such scans are repeatable.
Kudu Tablet Servers now properly retain historical data when a row with a given primary key is inserted and deleted, followed by the insertion of a new row with the same key. Previous versions of Kudu would not retain history in such situations. This allows the server to return correct results for snapshot scans with a timestamp in the past, even in the presence of such "reinsertion" scenarios.
The Kudu clients now automatically retain the timestamp of their latest
successful read or write operation. Scans using the
without a client-provided timestamp automatically assign a timestamp
higher than the timestamp of their most recent write. Writes also propagate
the timestamp, ensuring that sequences of operations with causal dependencies
between them are assigned increasing timestamps. Together, these changes
allow clients to achieve read-your-writes consistency, and also ensure
that snapshot scans performed by other clients return causally-consistent
Kudu servers now automatically limit the number of log files.
The number of log files retained can be configured using the
max_log_files flag. By default, 10 log files will be retained
at each severity level.
The logging in the Java and C++ clients has been substantially quieted. Clients no longer log messages in normal operation unless there is some kind of error.
The C++ client now includes a
API which can limit the amount of memory used to buffer
errors from asynchronous operations.
The Java client now fetches tablet locations from the Kudu Master in batches of 1000, increased from batches of 10 in prior versions. This can substantially improve the performance of Spark and Impala queries running against Kudu tables with large numbers of tablets.
Table metadata lock contention in the Kudu Master was substantially reduced. This improves the performance of tablet location lookups on large clusters with a high degree of concurrency.
Lock contention in the Kudu Tablet Server during high-concurrency write workloads was also reduced. This can reduce CPU consumption and improve performance when a large number of concurrent clients are writing to a smaller number of a servers.
Lock contention when writing log messages has been substantially reduced. This source of contention could cause high tail latencies on requests, and when under high load could contribute to cluster instability such as election storms and request timeouts.
BITSHUFFLE column encoding has been optimized to use the
instruction set present on processors including Intel® Sandy Bridge
and later. Scans on
BITSHUFFLE-encoded columns are now up to 30% faster.
kudu tool now accepts hyphens as an alternative to underscores
when specifying actions. For example,
kudu local-replica copy-from-remote
may be used as an alternative to
kudu local_replica copy_from_remote.
Fixed a long-standing issue in which running Kudu on
ext4 file systems
could cause file system corruption.
Implemented an LRU cache for open files, which prevents running out of
file descriptors on long-lived Kudu clusters. By default, Kudu will
limit its file descriptor usage to half of its configured
Fixed an issue which caused data corruption and crashes in the case that
a table had a non-composite (single-column) primary key, and that column
was specified to use
BITSHUFFLE encodings. If a
table with an affected schema was written in previous versions of Kudu,
the corruption will not be automatically repaired; users are encouraged
to re-insert such tables after upgrading to Kudu 1.2 or later.
Fixed a bug in the Spark
KuduRDD implementation which could cause
rows in the result set to be silently skipped in some cases.
KUDU-1551 Fixed an issue in which the tablet server would crash on restart in the case that it had previously crashed during the process of allocating a new WAL segment.
KUDU-1764 Fixed an issue where Kudu servers would leak approximately 16-32MB of disk space for every 10GB of data written to disk. After upgrading to Kudu 1.2 or later, any disk space leaked in previous versions will be automatically recovered on startup.
KUDU-1750 Fixed an issue where the API to drop a range partition would drop any partition with a matching lower or upper bound, rather than any partition with matching lower and upper bound.
Fixed an issue in the Java client where equality predicates which compared
an integer column to its maximum possible value (e.g.
would return incorrect results.
kudu-client Java artifact to properly shade classes in the
com.google.thirdparty namespace. The lack of proper shading in prior
releases could cause conflicts with certain versions of Google Guava.
Fixed shading issues in the
kudu-flume-sink Java artifact. The sink
now expects that Hadoop dependencies are provided by Flume, and properly
shades the Kudu client’s dependencies.
Fixed a few issues using the Python client library from Python 3.
Kudu 1.2.0 is wire-compatible with previous versions of Kudu:
Kudu 1.2 clients may connect to servers running Kudu 1.0. If the client uses features that are not available on the target server, an error will be returned.
Kudu 1.0 clients may connect to servers running Kudu 1.2 without limitations.
Rolling upgrade between Kudu 1.1 and Kudu 1.2 servers is believed to be possible though has not been sufficiently tested. Users are encouraged to shut down all nodes in the cluster, upgrade the software, and then restart the daemons on the new version.
The replication factor of tables is now limited to a maximum of 7. In addition, it is no longer allowed to create a table with an even replication factor.
GROUP_VARINT encoding is now deprecated. Kudu servers have never supported
this encoding, and now the client-side constant has been deprecated to match the
Kudu 1.2.0 introduces several new restrictions on schemas, cell size, and identifiers:
By default, Kudu will not permit the creation of tables with more than 300 columns. We recommend schema designs that use fewer columns for best performance.
No individual cell may be larger than 64KB. The cells making up a a composite key are limited to a total of 16KB after the internal composite-key encoding done by Kudu. Inserting rows not conforming to these limitations will result in errors being returned to the client.
Identifiers such as column and table names are now restricted to be valid UTF-8 strings. Additionally, a maximum length of 256 characters is enforced.
The Kudu 1.2 Java client is API- and ABI-compatible with Kudu 1.1. Applications written against Kudu 1.1 will compile and run against the Kudu 1.2 client and vice-versa.
The Kudu 1.2 C++ client is API- and ABI-forward-compatible with Kudu 1.1. Applications written and compiled against the Kudu 1.1 client will run without modification against the Kudu 1.2 client. Applications written and compiled against the Kudu 1.2 client will run without modification against the Kudu 1.1 client unless they use one of the following new APIs:
The Kudu 1.2 Python client is API-compatible with Kudu 1.1. Applications written against Kudu 1.1 will continue to run against the Kudu 1.2 client and vice-versa.
The Python client has been brought up to feature parity with the Java and C++ clients and as such the package version will be brought to 1.1 with this release (from 0.3). A list of the highlights can be found below.
Improved Partial Row semantics
Range partition support
Scan Token API
Enhanced predicate support
Support for all Kudu data types (including a mapping of Python’s
Alter table support
Enabled Read at Snapshot for Scanners
Enabled Scanner Replica Selection
A few bug fixes for Python 3 in addition to various other improvements.
IN LIST predicate pushdown support was added to allow optimized execution of filters which match on a set of column values. Support for Spark, Map Reduce and Impala queries utilizing IN LIST pushdown is not yet complete.
The Java client now features client-side request tracing in order to help troubleshoot timeouts. Error messages are now augmented with traces that show which servers were contacted before the timeout occurred instead of just the last error. The traces also contain RPCs that were required to fulfill the client’s request, such as contacting the master to discover a tablet’s location. Note that the traces are not available for successful requests and are not programmatically queryable.
Kudu now publishes JAR files for Spark 2.0 compiled with Scala 2.11 along with the existing Spark 1.6 JAR compiled with Scala 2.10.
The Java client now allows configuring scanners to read from the closest replica instead of
the known leader replica. The default remains the latter. Use the relevant
enum with the scanner’s builder to change this behavior.
Tablet servers use a new policy for retaining write-ahead log (WAL) segments. Previously, servers used the 'log_min_segments_to_retain' flag to prioritize any flushes which were retaining log segments past the configured value (default 2). This policy caused servers to flush in-memory data more frequently than necessary, limiting write performance.
The new policy introduces a new flag 'log_target_replay_size_mb' which determines the threshold at which write-ahead log retention will prioritize flushes. The new flag is considered experimental and users should not need to modify its value.
The improved policy has been seen to improve write performance in some use cases by a factor of 2x relative to the old policy.
Kudu’s implementation of the Raft consensus algorithm has been improved to include a "pre-election" phase. This can improve the stability of tablet leader election in high-load scenarios, especially if each server hosts a high number of tablets.
Tablet server start-up time has been substantially improved in the case that the server contains a high number of tombstoned tablet replicas.
kudu tablet leader_step_down has been added to manually force a leader to step down.
kudu remote_replica copy has been added to manually copy a replica from
one running tablet server to another.
kudu local_replica delete has been added to delete a replica of a tablet.
kudu test loadgen tool has been added to replace the obsoleted
insert-generated-rows standalone binary. The new tool is enriched with
additional functionality and can be used to run load generation tests against
a Kudu cluster.
Kudu 1.1.0 is wire-compatible with previous versions of Kudu:
Kudu 1.1 clients may connect to servers running Kudu 1.0. If the client uses the new 'IN LIST' predicate type, an error will be returned.
Kudu 1.0 clients may connect to servers running Kudu 1.1 without limitations.
Rolling upgrade between Kudu 1.0 and Kudu 1.1 servers is believed to be possible though has not been sufficiently tested. Users are encouraged to shut down all nodes in the cluster, upgrade the software, and then restart the daemons on the new version.
The C++ client no longer requires the old gcc5 ABI. Which ABI is actually used depends on the compiler configuration. Some new distros (e.g. Ubuntu 16.04) will use the new ABI. Your application must use the same ABI as is used by the client library; an easy way to guarantee this is to use the same compiler to build both.
The C++ client’s
KuduSession::CountBufferedOperations() method is
deprecated. Its behavior is inconsistent unless the session runs in the
MANUAL_FLUSH mode. Instead, to get number of buffered operations, count
invocations of the
KuduSession::Apply() method since last
KuduSession::Flush() call or, if using asynchronous flushing, since last
invocation of the callback passed into
The Java client’s
OperationResponse.getWriteTimestamp method was renamed to
to emphasize that it doesn’t return milliseconds, unlike what its Javadoc indicated. The renamed
method was also hidden from the public APIs and should not be used.
The Java client’s sync API (
KuduScanner) used to throw either
NonRecoverableException or a
TimeoutException for a timeout, and now it’s only possible for the
client to throw the former.
The Java client’s handling of errors in
KuduSession was modified so that subclasses of
KuduException are converted into RowErrors instead of being thrown.
Apache Kudu 1.0.1 is a bug fix release, with no new features or backwards incompatible changes.
KUDU-1681 Fixed a bug in the tablet server which could cause a crash when the DNS lookup during master heartbeat failed.
KUDU-1660: Fixed a bug which would cause the Kudu master and tablet server to fail to start on single CPU systems.
KUDU-1652: Fixed a bug
that would cause the C++ client, tablet server, and Java client to crash or
throw an exception when attempting to scan a table with a predicate which
IS NOT NULL on a non-nullable column. For instance, setting a
⇐ 127 predicate on an
INT8 column could trigger this bug, since the
predicate only filters null values.
KUDU-1651: Fixed a bug that would cause the tablet server to crash when evaluating a scan with predicates over a dictionary encoded column containing an entire block of null values.
KUDU-1623: Fixed a bug that would cause the tablet server to crash when handling UPSERT operations that only set values for the primary key columns.
Gerrit #4488 Fixed a bug in the Java client’s KuduException class which could cause an unexpected NullPointerException to be thrown when the exception did not have an associated message.
KUDU-1090 Fixed a bug in the memory tracker which could cause a rare crash during tablet server startup.
After approximately a year of beta releases, Apache Kudu has reached version 1.0. This version number signifies that the development team feels that Kudu is stable enough for usage in production environments.
If you are new to Kudu, check out its list of features and benefits.
Kudu 1.0.0 delivers a number of new features, bug fixes, and optimizations.
Removal of multiversion concurrency control (MVCC) history is now supported. This is known as tablet history GC. This allows Kudu to reclaim disk space, where previously Kudu would keep a full history of all changes made to a given table since the beginning of time. Previously, the only way to reclaim disk space was to drop a table.
Kudu will still keep historical data, and the amount of history retained is
controlled by setting the configuration flag
which defaults to 15 minutes (expressed in seconds). The timestamp
represented by the current time minus
tablet_history_max_age_sec is known
as the ancient history mark (AHM). When a compaction or flush occurs, Kudu
will remove the history of changes made prior to the ancient history mark.
This only affects historical data; currently-visible data will not be
removed. A specialized maintenance manager background task to remove existing
"cold" historical data that is not in a row affected by the normal compaction
process will be added in a future release.
Most of Kudu’s command line tools have been consolidated under a new
kudu tool. This reduces the number of large binaries distributed
with Kudu and also includes much-improved help output.
The Kudu Flume Sink now supports processing events containing Avro-encoded
records, using the new
Administrative tools including
kudu cluster ksck now support running
against multi-master Kudu clusters.
The output of the
ksck tool is now colorized and much easier to read.
The C++ client API now supports writing data in
This can provide higher throughput for ingest workloads.
The performance of comparison predicates on dictionary-encoded columns has been substantially optimized. Users are encouraged to use dictionary encoding on any string or binary columns with low cardinality, especially if these columns will be filtered with predicates.
The Java client is now able to prune partitions from scanners based on the provided predicates. For example, an equality predicate on a hash-partitioned column will now only access those tablets that could possibly contain matching data. This is expected to improve performance for the Spark integration as well as applications using the Java client API.
The performance of compaction selection in the tablet server has been substantially improved. This can increase the efficiency of the background maintenance threads and improve overall throughput of heavy write workloads.
The policy by which the tablet server retains write-ahead log (WAL) files has been improved so that it takes into account other replicas of the tablet. This should help mitigate the spurious eviction of tablet replicas on machines that temporarily lag behind the other replicas.
Kudu 1.0.0 maintains client-server wire-compatibility with previous releases. Applications using the Kudu client libraries may be upgraded either before, at the same time, or after the Kudu servers.
Kudu 1.0.0 does not maintain server-server wire compatibility with previous releases. Therefore, rolling upgrades between earlier versions of Kudu and Kudu 1.0.0 are not supported.
kudu-pbc-dump tool has been removed. The same functionality is now
kudu pbc dump.
kudu-ksck tool has been removed. The same functionality is now
kudu cluster ksck.
cfile-dump tool has been removed. The same functionality is now
kudu fs cfile dump.
log-dump tool has been removed. The same functionality is now
kudu wal dump and
kudu local_replica dump wals.
kudu-admin tool has been removed. The same functionality is now
kudu table and
kudu-fs_dump tool has been removed. The same functionality is now
kudu fs dump.
kudu-ts-cli tool has been removed. The same functionality is now
kudu remote_replica, and
kudu-fs_list tool has been removed and some similar useful
functionality has been moved under 'kudu local_replica'.
Some configuration flags are now marked as 'unsafe' and 'experimental'. Such flags
are disallowed by default. Users may access these flags by enabling the additional
--unlock_experimental_flags. Usage of such flags
is not recommended, as the flags may be removed or modified with no deprecation period
and without notice in future Kudu releases.
TIMESTAMP column type has been renamed to
UNIXTIME_MICROS in order to
reduce confusion between Kudu’s timestamp support and the timestamps supported
by other systems such as Apache Hive and Apache Impala (incubating). Existing
tables will automatically be updated to use the new name for the type.
Clients upgrading to the new client libraries must move to the new name for the type. Clients using old client libraries will continue to operate using the old type name, even when connected to clusters that have been upgraded. Similarly, if clients are upgraded before servers, existing timestamp columns will be available using the new type name.
KuduSession methods in the C++ library are no longer advertised as thread-safe
to have one set of semantics for both C++ and Java Kudu client libraries.
KuduScanToken::TabletServers method in the C++ library has been removed.
The same information can now be found in the KuduScanToken::tablet method.
KuduEventProducer interface used to process Flume events into Kudu operations
for the Kudu Flume Sink has changed, and has been renamed
The existing `KuduEventProducer`s have been updated for the new interface, and have
been renamed similarly.
Kudu 0.10.0 delivers a number of new features, bug fixes, and optimizations, detailed below.
Kudu 0.10.0 maintains wire-compatibility with previous releases, meaning that applications using the Kudu client libraries may be upgraded either before, at the same time, or after the Kudu servers. However, if you begin using new features of Kudu 0.10.0 such as manually range-partitioned tables, you must first upgrade all clients to this release.
This release does not maintain full Java API or ABI compatibility with Kudu 0.9.x due to a package rename and some other small changes. See below for details.
Gerrit #3737 The Java client has been repackaged
org.apache.kudu instead of
org.kududb. Import statements for Kudu classes must
be modified in order to compile against 0.10.0. Wire compatibility is maintained.
Gerrit #3055 The Java client’s
synchronous API methods now throw
KuduException instead of
Existing code that catches
Exception should still compile, but introspection of an
exception’s message may be impacted. This change was made to allow thrown exceptions to be
queried more easily using
KuduException.getStatus and calling one of
For example, an operation that tries to delete a table that doesn’t exist would return a
`Status that returns true when queried on
The Java client’s
KuduTable.getTabletsLocations set of methods is now
deprecated. Additionally, they now take an exclusive end partition key instead
of an inclusive key. Applications are encouraged to use the scan tokens API
instead of these methods in the future.
The C++ API for specifying split points on range-partitioned tables has been improved to make it easier for callers to properly manage the ownership of the provided rows.
TableCreator::split_rows API took a
vector<const KuduPartialRow*>, which
made it very difficult for the calling application to do proper error handling with
cleanup when setting the fields of the
KuduPartialRow. This API has been now been
deprecated and replaced by a new method
TableCreator::add_range_split which allows
easier use of smart pointers for safe memory management.
The Java client’s internal buffering has been reworked. Previously, the number of buffered write operations was constrained on a per-tablet-server basis. Now, the configured maximum buffer size constrains the total number of buffered operations across all tablet servers in the cluster. This provides a more consistent bound on the memory usage of the client regardless of the size of the cluster to which it is writing.
This change can negatively affect the write performance of Java clients which rely on
buffered writes. Consider using the
setMutationBufferSpace API to increase a
session’s maximum buffer size if write performance seems to be degraded after upgrading
to Kudu 0.10.0.
The "remote bootstrap" process used to copy a tablet replica from one host to another has been renamed to "Tablet Copy". This resulted in the renaming of several RPC metrics. Any users previously explicitly fetching or monitoring metrics related to Remote Bootstrap should update their scripts to reflect the new names.
The SparkSQL datasource for Kudu no longer supports mode
Overwrite. Users should
use the new
KuduContext.upsertRows method instead. Additionally, inserts using the
datasource are now upserts by default. The older behavior can be restored by setting
operation parameter to
Users may now manually manage the partitioning of a range-partitioned table. When a table is created, the user may specify a set of range partitions that do not cover the entire available key space. A user may add or drop range partitions to existing tables.
This feature can be particularly helpful with time series workloads in which new partitions can be created on an hourly or daily basis. Old partitions may be efficiently dropped if the application does not need to retain historical data past a certain point.
This feature is considered experimental for the 0.10 release. More details of the new feature can be found in the accompanying blog post.
Support for running Kudu clusters with multiple masters has been stabilized. Users may start a cluster with three or five masters to provide fault tolerance despite a failure of one or two masters, respectively.
Note that certain tools (e.g.
ksck) are still lacking complete support for
multiple masters. These deficiencies will be addressed in a following release.
Kudu now supports the ability to reserve a certain amount of free disk space in each of its configured data directories. If a directory’s free disk space drops to less than the configured minimum, Kudu will stop writing to that directory until space becomes available. If no space is available in any configured directory, Kudu will abort.
This feature may be configured using the
The Spark integration’s
KuduContext now supports four new methods for writing to
deleteRows. These are
now the preferred way to write to Kudu tables from Spark.
has been improved and now detects problems such as when a tablet does not have
a majority of replicas on live tablet servers, or if those replicas aren’t in a
good state. Users who currently depend on the tool to detect inconsistencies may now see
failures when before they wouldn’t see any.
Gerrit #3477 The way operations are buffered in the Java client has been reworked. Previously, the session’s buffer size was set per tablet, meaning that a buffer size of 1,000 for 10 tablets being written to allowed for 10,000 operations to be buffered at the same time. With this change, all the tablets share one buffer, so users might need to set a bigger buffer size in order to reach the same level of performance as before.
Gerrit #3674 Added LESS and GREATER options for column predicates.
KUDU-1444 added support for passing
back basic per-scan metrics (e.g cache hit rate) from the server to the C++ client. See the
KuduScanner::GetResourceMetrics() API for detailed usage. This feature will be supported
in the Java client API in a future release.
KUDU-1446 improved the order in which the tablet server evaluates predicates, so that predicates on smaller columns are evaluated first. This may improve performance on queries which apply predicates on multiple columns of different sizes.
KUDU-1398 improved the storage efficiency of Kudu’s internal primary key indexes. This optimization should decrease space usage and improve random access performance, particularly for workloads with lengthy primary keys.
Gerrit #3541 Fixed a problem in the Java client
whereby an RPC could be dropped when a connection to a tablet server or master was forcefully
closed on the server-side while RPCs to that server were in the process of being encoded.
The effect was that the RPC would not be sent, and users of the synchronous API would receive
TimeoutException. Several other Java client bugs which could cause similar spurious timeouts
were also fixed in this release.
Gerrit #3724 Fixed a problem in the Java client whereby an RPC could be dropped when a socket timeout was fired while that RPC was being sent to a tablet server or master. This would manifest itself in the same way Gerrit #3541.
KUDU-1538 fixed a bug in which recycled block identifiers could cause the tablet server to lose data. Following this bug fix, block identifiers will no longer be reused.
This is the first release of Apache Kudu as a top-level (non-incubating) project!
The default false positive rate for Bloom filters has been changed from 1% to 0.01%. This will increase the space consumption of Bloom filters by a factor of two (from approximately 10 bits per row to approximately 20 bits per row). This is expected to substantially improve the performance of random-write workloads at the cost of an incremental increase in disk space usage.
The Kudu C++ client library now has Doxygen-based API documentation available online.
Kudu now uses the Raft consensus algorithm even for unreplicated tables. This change simplifies code and will also allow administrators to enable replication on a previously-unreplicated table. This change is internal and should not be visible to users.
Kudu 0.9.1 delivers incremental bug fixes over Kudu 0.9.0. It is fully compatible with Kudu 0.9.0.
KUDU-1469 fixed a bug in our Raft consensus implementation that could cause a tablet to stop making progress after a leader election.
Gerrit #3456 fixed a bug in which servers under high load could store metric information in incorrect memory locations, causing crashes or data corruption.
Gerrit #3457 fixed a bug in which errors from the Java client would carry an incorrect error message.
Several other small bug fixes were backported to improve stability.
Kudu 0.9.0 delivers incremental features, improvements, and bug fixes over the previous versions.
KuduTableInputFormat command has changed the way in which it handles
scan predicates, including how it serializes predicates to the job configuration
object. The new configuration key is
TableInputFormatConfigurator are not affected.
kudu-spark sub-project has been renamed to follow naming conventions for
Scala. The new name is
Default table partitioning has been removed. All tables must now be created with explicit partitioning. Existing tables are unaffected. See the schema design guide for more details.
KUDU-1002 Added support for
UPSERT operations, whereby a row is inserted if it does not already exist, but
updated if it does. Support for
UPSERT is included in Java, C++, and Python APIs,
but not in Impala.
KUDU-1306 Scan token API for creating partition-aware scan descriptors. This API simplifies executing parallel scans for clients and query engines.
Gerrit 2848 Added a kudu datasource
for Spark. This datasource uses the Kudu client directly instead of
using the MapReduce API. Predicate pushdowns for
spark-sql and Spark filters are
included, as well as parallel retrieval for multiple tablets and column projections.
See an example of Kudu integration with Spark.
Gerrit 2992 Added the ability to update and insert from Spark using a Kudu datasource.
KUDU-678 Fixed a leak that happened during DiskRowSet compactions where tiny blocks were still written to disk even if there were no REDO records. With the default block manager, it usually resulted in block containers with thousands of tiny blocks.
KUDU-1437 Fixed a data corruption issue that occured after compacting sequences of negative INT32 values in a column that was configured with RLE encoding.
All Kudu clients have longer default timeout values, as listed below.
The default operation timeout and the default admin operation timeout are now set to 30 seconds instead of 10.
The default socket read timeout is now 10 seconds instead of 5.
The default admin timeout is now 30 seconds instead of 10.
The default RPC timeout is now 10 seconds instead of 5.
The default scan timeout is now 30 seconds instead of 15.
Some default settings related to I/O behavior during flushes and compactions have been changed:
The default for
flush_threshold_mb has been increased from 64MB to 1000MB. The default
cfile_do_on_finish has been changed from
Experiments using YCSB indicate that these
values will provide better throughput for write-heavy applications on typical server hardware.
Kudu 0.8.0 delivers incremental features, improvements, and bug fixes over the previous versions.
To upgrade to Kudu 0.8.0, see Upgrade from 0.7.1 to 0.8.0.
0.8.0 clients are not fully compatible with servers running Kudu 0.7.1 or lower. In particular, scans that specify column predicates will fail. To work around this issue, upgrade all Kudu servers before upgrading clients.
KUDU-839 Java RowError now uses an enum error code.
Gerrit 2138 The handling of column predicates has been re-implemented in the server and clients.
KUDU-1379 Partition pruning has been implemented for C++ clients (but not yet for the Java client). This feature allows you to avoid reading a tablet if you know it does not serve the row keys you are querying.
Gerrit 2641 Kudu now uses
earliest-deadline-first RPC scheduling and rejection. This changes the behavior
of the RPC service queue to prevent unfairness when processing a backlog of RPC
threads and to increase the likelihood that an RPC will be processed before it
can time out.
KUDU-1337 Tablets from tables that were deleted might be unnecessarily re-bootstrapped when the leader gets the notification to delete itself after the replicas do.
KUDU-969 If a tablet server shuts down while compacting a rowset and receiving updates for it, it might immediately crash upon restart while bootstrapping that rowset’s tablet.
KUDU-1354 Due to a bug in Kudu’s MVCC implementation where row locks were released before the MVCC commit happened, flushed data would include out-of-order transactions, triggering a crash on the next compaction.
KUDU-1322 The C++ client now retries write operations if the tablet it is trying to reach has already been deleted.
Gerrit 2571 Due to a bug in the
Java client, users were unable to close the
kudu-spark shell because of
lingering non-daemon threads.
Gerrit 2239 The concept of "feature flags" was introduced in order to manage compatibility between different Kudu versions. One case where this is helpful is if a newer client attempts to use a feature unsupported by the currently-running tablet server. Rather than receiving a cryptic error, the user gets an error message that is easier to interpret. This is an internal change for Kudu system developers and requires no action by users of the clients or API.
Kudu 0.7.1 is a bug fix release for 0.7.0.
KUDU-1325 fixes a tablet server crash that could occur during table deletion. In some cases, while a table was being deleted, other replicas would attempt to re-replicate tablets to servers that had already processed the deletion. This could trigger a race condition that caused a crash.
KUDU-1341 fixes a potential data corruption and crash that could happen shortly after tablet server restarts in workloads that repeatedly delete and re-insert rows with the same primary key. In most cases, this corruption affected only a single replica and could be repaired by re-replicating from another.
KUDU-1343 fixes a bug in the Java client that occurs when a scanner has to scan multiple batches from one tablet and then start scanning from another. In particular, this would affect any scans using the Java client that read large numbers of rows from multi-tablet tables.
KUDU-1345 fixes a bug where in some cases the hybrid clock could jump backwards, resulting in a crash followed by an inability to restart the affected tablet server.
KUDU-1360 fixes a bug in the kudu-spark module
which prevented reading rows with
Kudu 0.7.0 is the first release done as part of the Apache Incubator and includes a number of changes, new features, improvements, and fixes.
The upgrade instructions can be found at Upgrade from 0.6.0 to 0.7.0.
The C++ client includes a new API,
KuduScanBatch, which performs better when a
large number of small rows are returned in a batch. The old API of
|This change is API-compatible but not ABI-compatible.|
The default replication factor has been changed from 1 to 3. Existing tables will
continue to use the replication factor they were created with. Applications that create
tables may not work properly if they assume a replication factor of 1 and fewer than
3 replicas are available. To use the previous default replication factor, start the
master with the configuration flag
The Python client has been completely rewritten, with a focus on improving code quality and testing. The read path (scanners) has been improved by adding many of the features already supported by the C++ and Java clients. The Python client is no longer considered experimental.
With the goal of Spark integration in mind, a new
kuduRDD API has been added,
newAPIHadoopRDD and includes a default source for Spark SQL.
The Java client includes new methods
KuduSession. These methods allow you to count and
retrieve outstanding row errors when configuring sessions with
New server-level metrics allow you to monitor CPU usage and context switching.
Kudu now builds on RHEL 7, CentOS 7, and SLES 12. Extra instructions are included for SLES 12.
The file block manager’s performance was improved, but it is still not recommended for real-world use.
The master now attempts to spread tablets more evenly across the cluster during table creation. This has no impact on existing tables, but will improve the speed at which under-replicated tablets are re-replicated after a tablet server failure.
All licensing documents have been modified to adhere to ASF guidelines.
Kudu now requires an out-of-tree build directory. Review the build instructions for additional information.
C` client library is now explicitly built against the
link:https://gcc.gnu.org/onlinedocs/libstdc/manual/using_dual_abi.html[old gcc5 ABI].
If you use gcc5 to build a Kudu application, your application must use the old ABI
as well. This is typically achieved by defining the `_GLIBCXX_USE_CXX11_ABI macro
at compile-time when building your application. For more information, see the
previous link and link:http://developerblog.redhat.com/2015/02/05/gcc5-and-the-c11-abi/.
The Python client is no longer considered experimental.
The 0.6.0 release contains incremental improvements and bug fixes. The most notable changes are:
The Java client’s CreateTableBuilder and AlterTableBuilder classes have been renamed
to CreateTableOptions and AlterTableOptions. Their methods now also return
allowing them to be used as builders.
The Java client’s AbstractKuduScannerBuilder#maxNumBytes() setter is now called batchSizeBytes as is the corresponding property in AsyncKuduScanner. This makes it consistent with the C++ client.
The "kudu-admin" tool can now list and delete tables via its new subcommands "list_tables" and "delete_table <table_name>".
OSX is now supported for single-host development. Please consult its specific installation instructions in OS X.
Kudu 0.5.0 was the first public release. As such, no improvements or changes were noted in its release notes.