|This document applies to Apache Kudu version 1.15.0. Please consult the documentation of the appropriate release that’s applicable to the version of the Kudu cluster.|
The primary key may not be changed after the table is created. You must drop and recreate a table to select a new primary key.
The columns which make up the primary key must be listed first in the schema.
The primary key of a row may not be modified using the
To modify a row’s primary key, the row must be deleted and re-inserted with
the modified key. Such a modification is non-atomic.
BOOL types are not allowed as part of a
primary key definition. Additionally, all columns that are part of a primary
key definition must be
Auto-generated primary keys are not supported.
Cells making up a composite primary key are limited to a total of 16KiB after the internal composite-key encoding done by Kudu.
CHAR and complex types like
STRUCT are not yet
Type, nullability and type attributes (i.e. precision and scale of
VARCHAR) of existing columns cannot be changed by altering the
Tables can have a maximum of 300 columns by default.
Tables must have an odd number of replicas, with a maximum of 7.
The replication factor is set at table creation time and cannot be changed later. As a workaround, create a new table with the same schema and the desired replication factor and insert the contents of the old table into the new one.
By default, cells cannot be larger than 64KiB before encoding or compression.
Secondary indexes are not supported.
Multi-row transactions are not yet supported.
Relational features, like foreign keys, are not supported.
Identifiers such as column and table names are restricted to be valid UTF-8 strings. Additionally, a maximum length of 256 characters is enforced.
Dropping a column does not immediately reclaim space. Compaction must run first.
There is no way to run compaction manually, but dropping the table will reclaim the space immediately.
Tables must be manually pre-split into tablets using simple or compound primary keys. Automatic splitting is not yet possible. Range partitions may be added or dropped after a table has been created. See Schema Design for more information.
Data in existing tables cannot currently be automatically repartitioned. As a workaround, create a new table with the new partitioning and insert the contents of the old table.
Tablets that lose a majority of replicas (such as 1 left out of 3) require manual intervention to be repaired.
Recommended maximum point-to-point latency within a Kudu cluster is 20 milliseconds.
Recommended minimum point-to-point bandwidth within a Kudu cluster is 10 Gbps.
If you intend to use the location awareness feature to place tablet servers in different locations, it is recommended to measure the bandwidth and latency between servers to ensure they fit within the above guidelines.
All masters must be started at the same time when the cluster is started for the very first time.
Production deployments should configure a least 4 GiB of memory for tablet servers, and ideally more than 16 GiB when approaching the data and tablet Scale limits.
Write ahead logs (WAL) can only be stored on one disk.
Tablet servers can’t change address/port.
Kudu has a hard requirement on having up-to-date NTP. Kudu masters and tablet servers will crash when out of sync.
Kudu is known to run seamlessly across a wide array of environments and workloads with minimal expertise and configuration at the following scale:
3 master servers
100 tablet servers
8 TiB of stored data per tablet server, post-replication and post-compression.
1000 tablets per tablet server, post-replication.
60 tablets per table, per tablet server, at table-creation time.
10 GiB of stored data per tablet.
Staying within these limits will provide the most predictable and straightforward Kudu experience.
However, experienced users who run on modern hardware, use the latest versions of Kudu, test and tune Kudu for their use case, and work closely with the community, can achieve much higher scales comfortably. Below are some anecdotal values that have been seen in real world production clusters:
3 master servers
300+ tablet servers
10+ TiB of stored data per tablet server, post-replication and post-compression.
4000+ tablets per tablet server, post-replication.
50 GiB of stored data per tablet. Going beyond this can cause issues such a reduced performance, compaction issues, and slow tablet startup times.
Row-level authorization is not available.
Data encryption at rest is not directly built into Kudu. Encryption of
Kudu data at rest can be achieved through the use of local block device
encryption software such as
Kudu server Kerberos principals must follow the pattern
Configuring an alternate Kerberos principal is not supported.
The highest supported version of the TLS protocol is TLSv1.2
The following are known bugs and issues with the current release of Kudu. They will be addressed in later releases. Note that this list is not exhaustive, and is meant to communicate only the most important known issues.
If a tablet server has a very large number of tablets, it may take several minutes to start up. It is recommended to limit the number of tablets per server to 1000 or fewer. Consider this limitation when pre-splitting your tables. If you notice slow start-up times, you can monitor the number of tablets per server in the web UI.
NVM-based cache doesn’t work reliably on RH6/CentOS6 (see KUDU-2978).
When upgrading a Kudu cluster to 1.11.0 version with existing pre-1.11.0
on_disk_size metrics might produce
inconsistent readings in some scenarios
In Kudu 1.10.0 and Kudu 1.11.0, the kudu-binary JAR (targeted for
containerized Kudu deployments using mini-cluster) contains libnuma dynamic
library. Also, if building Kudu binaries in release mode with default cmake
settings, the libnuma library is linked statically with the Kudu binaries
-DKUDU_LINK=dynamic when running cmake to avoid that). The library is
licensed under LGPL v.2.1, however the ASF thirdparty license policy
explicitly prohibits including such contents into releases: see
Category X. This
issue has been addressed in 1.10.1 and 1.11.1 patch releases correspondingly
Due to a bug in SSSD PAC plugin of version prior to 1.16, a KRPC connection negotiation may stuck and the whole process wouldn’t be able to negotiate any new connection for about 5 minutes in a secure Kudu cluster using SSSD. If using SSSD in secure Kudu cluster deployments, make sure SSSD packages are of version 1.16 or newer (see KUDU-3217).