Apache Kudu Scaling Guide

This document describes in detail how Kudu scales with respect to various system resources, including memory, file descriptors, and threads. See the scaling limits for the maximum recommended parameters of a Kudu cluster. They can be used to estimate roughly the number of servers required for a given quantity of data.

The recommendations and conclusions here are only approximations. Appropriate numbers depend on use case. There is no substitute for measurement and monitoring of resources used during a representative workload.

Terms

We will use the following terms:

hot replica: A tablet replica that is continuously receiving writes. For example, in a time series use case, tablet replicas for the most recent range partition on a time column would be continuously receiving the latest data, and would be hot replicas.
cold replica: A tablet replica that is not hot, i.e. a replica that is not frequently receiving writes, for example, once every few minutes. A cold replica may be read from. For example, in a time series use case, tablet replicas for previous range partitions on a time column would not receive writes at all, or only occasionally receive late updates or additions, but may be constantly read.
data on disk: The total amount of data stored on a tablet server across all disks, post-replication, post-compression, and post-encoding.

Example Workload

The sections below perform sample calculations using the following parameters:

200 hot replicas per tablet server
1600 cold replicas per tablet server
8TB of data on disk per tablet server (about 4.5GB/replica)
512MB block cache
40 cores per server
limit of 32000 file descriptors per server
a read workload with 1 frequently-scanned table with 40 columns

This workload resembles a time series use case, where the hot replicas correspond to the most recent range partition on time.

Memory

The flag --memory_limit_hard_bytes determines the maximum amount of memory that a Kudu tablet server may use. The amount of memory used by a tablet server scales with data size, write workload, and read concurrency. The following table provides numbers that can be used to compute a rough estimate of memory usage.

Table 1. Tablet Server Memory Usage
Type	Multiplier	Description
Memory required per TB of data on disk	1.5GB per 1TB data on disk	Amount of memory per unit of data on disk required for basic operation of the tablet server.
Hot Replicas' MemRowSets and DeltaMemStores	minimum 128MB per hot replica	Minimum amount of data to flush per MemRowSet flush. For most use cases, updates should be rare compared to inserts, so the DeltaMemStores should be very small.
Scans	256KB per column per core for read-heavy tables	Amount of memory used by scanners, and which will be constantly needed for tables which are constantly read.
Block Cache	Fixed by `--block_cache_capacity_mb` (default 512MB)	Amount of memory reserved for use by the block cache.

Using this information for the example load gives the following breakdown of memory usage:

Table 2. Example Tablet Server Memory Usage
Type	Amount
8TB data on disk	8TB * 1.5GB / 1TB = 12GB
200 hot replicas	200 * 128MB = 25.6GB
1 40-column, frequently-scanned table	40 * 40 * 256KB = 409.6MB
Block Cache	`--block_cache_capacity_mb=512` = 512MB
Expected memory usage	38.5GB
Recommended hard limit	52GB

Using this as a rough estimate of Kudu’s memory usage, select a memory limit so that the expected memory usage of Kudu is around 50-75% of the hard limit.

Verifying if a Memory Limit is sufficient

After configuring an appropriate memory limit with --memory_limit_hard_bytes, run a workload and monitor the Kudu tablet server process’s RAM usage. The memory usage should stay around 50-75% of the hard limit, with occasional spikes above 75% but below 100%. If the tablet server runs above 75% consistently, the memory limit should be increased.

Additionally, it’s also useful to monitor the logs for memory rejections, which look like:

Service unavailable: Soft memory limit exceeded (at 96.35% of capacity)

and watch the memory rejections metrics:

leader_memory_pressure_rejections
follower_memory_pressure_rejections
transaction_memory_pressure_rejections

Occasional rejections due to memory pressure are fine and act as backpressure to clients. Clients will transparently retry operations. However, no operations should time out.

File Descriptors

Processes are allotted a maximum number of open file descriptors (also referred to as fds). If a tablet server attempts to open too many fds, it will typically crash with a message saying something like "too many open files". The following table summarizes the sources of file descriptor usage in a Kudu tablet server process:

Table 3. Tablet Server File Descriptor Usage
Type	Multiplier	Description
File cache	Fixed by `--block_manager_max_open_files` (default 40% of process maximum)	Maximum allowed open fds reserved for use by the file cache.
Hot replicas	2 per WAL segment, 1 per WAL index	Number of fds used by hot replicas. See below for more explanation.
Cold replicas	3 per cold replica	Number of fds used per cold replica: 2 for the single WAL segment and 1 for the single WAL index.

Every replica has at least one WAL segment and at least one WAL index, and should have the same number of segments and indices; however, the number of segments and indices can be greater for a replica if one of its peer replicas is falling behind. WAL segment and index fds are closed as WALs are garbage collected.

Using this information for the example load gives the following breakdown of file descriptor usage, under the assumption that some replicas are lagging and using 10 WAL segments:

Table 4. Example Tablet Server File Descriptor Usage
Type	Amount
file cache	40% * 32000 fds = 12800 fds
1600 cold replicas	1600 cold replicas * 3 fds / cold replica = 4800 fds
200 hot replicas	(2 / segment * 10 segments/hot replica * 200 hot replicas) + (1 / index * 10 indices / hot replica * 200 hot replicas) = 6000 fds
Total	23600 fds

So for this example, the tablet server process has about 32000 - 23600 = 8400 fds to spare.

There is typically no downside to configuring a higher file descriptor limit if approaching the currently configured limit.

Threads

Processes are allotted a maximum number of threads by the operating system, and this limit is typically difficult or impossible to change. Therefore, this section is more informational than advisory.

If a Kudu tablet server’s thread count exceeds the OS limit, it will crash, usually with a message in the logs like "pthread_create failed: Resource temporarily unavailable". If the system thread count limit is exceeded, other processes on the same node may also crash.

Threads and threadpools are used all over Kudu for various purposes, but the number of threads found in nearly all of these does not scale with load or data/tablet size; instead, the number of threads is either a hardcoded constant, a constant defined by a configuration parameter, or based on a static dimension (such as the number of CPU cores).

The only exception to this is the WAL append thread, one of which exists for every "hot" replica. Note that all replicas may be considered hot at startup, so tablet servers' thread usage will generally peak when started and settle down thereafter.