Tales From A Lazy Fat DBA

Hi Fellas,
Back and this time with some performance tuning scopes for Cassandra DB during the ‘Compaction’ process.
Before i proceed, would like to explain a bit about the compaction in Cassandra and what exactly is this and why a necessary evil …

Compaction in Cassandra refers to the operation of merging multiple SSTables into a single new one. Typically, compaction is done in a database for two primary reasons:

– To reduce the storage usage.
– To improve read performance by merging keys and obtaining a consolidated index.

For example, in Apache Cassandra, data files are merged periodically to form compacted SSTables.

There is a good chance of contention happening in database due to Compaction activity as the Compaction increases I/O contention on SSTable data read. Writing data in Cassandra database is generally fast and the write impacts may not be seen but reading data from SSTables will be slow in case when I/O contention increases due to compaction activities and degrades the performance of the database.

First would like to discuss how to identify the compaction related contentions on the database.
– We can use the “nodetool tablestats” or the old “nodetool cfstats” command to
monitor or watch-keep SSTables.
Below is a sample result from one of the Cassandra database server, here we need to check
– Check if the count is keep on growing, because that points out that there may be contention between reading SST
and the compaction process.
– Read generally slows down due to an obvious reason of data distributed or fragmented across many SSTs and
Compaction running continuous in the background.

%nodetool tablestats -H dixit.playlist
Keyspace: dixit
Read Count: 182849
Read Latency: 0.11363755339104945 ms.
Write Count: 435355
Write Latency: 0.01956930550929701 ms.
Pending Flushes: 0
Table: standard1
SSTable count: 2
Space used (live): 51.62 MB
Space used (total): 51.62 MB
Space used by snapshots (total): 0 bytes
Off heap memory used (total): 302.36 KB
SSTable Compression Ratio: 0.0
Number of keys (estimate): 376390
Memtable cell count: 200120
Memtable data size: 45.16 MB
Memtable off heap memory used: 0 bytes
Memtable switch count: 2
Local read count: 182849
Local read latency: 0.125 ms
Local write count: 435355
Local write latency: 0.022 ms
Pending flushes: 0
Bloom filter false positives: 11
Bloom filter false ratio: 0.00000
Bloom filter space used: 265.81 KB
Bloom filter off heap memory used: 265.8 KB
Index summary off heap memory used: 36.57 KB
Compression metadata off heap memory used: 0 bytes
Compacted partition minimum bytes: 216 bytes
Compacted partition maximum bytes: 258 bytes
Compacted partition mean bytes: 258 bytes
Average live cells per slice (last five minutes): 1.0
Maximum live cells per slice (last five minutes): 1
Average tombstones per slice (last five minutes): 1.0
Maximum tombstones per slice (last five minutes): 1

Below is the command that can be used to check for compaction statistics, here you need to look at the ‘pending tasks’, and ‘bytes total in progress’.

$ nodetool compactionstats
pending tasks: 5
compaction type keyspace table completed total unit progress
Compaction Keyspace1 Standard1 282310680 302170540 bytes 93.43%
Compaction Keyspace1 Standard1 58457931 307520780 bytes 19.01%
Active compaction remaining time : 0h00m16s

Solution to the problem
1. First one is quite simple – Avoid merging of update/delete requests.
2. Reduce the frequency of in-memory objects (In Memtables) flush.

This can be done by increasing the size of the memtables to avoid or stop database to perform frequent flushes.
– Less number of flushes leads to fewer SSTs compaction.
– Less Compaction reduces the I/Contentions and this in turn improve reads.
– There are couple of parameters that you can adjust in your cassandra.yaml file to control the flushing.
i.e. memtable_flush_after_mins, memtable_throughput_in_mb , memtable_operations_in_millions.

3. One more solution but that only applies on systems where this stress in IO is not much frequent, we can reduce
the “thread priority” which reduces the IOs.
As lowering the priority slows down the compaction writes but only applies if it doesn’t happen frequently.

Add below lines in cassandra-env.conf file (Under /conf folder) to lower the compaction priority.

JVM_OPTS=”$JVM_OPTS -XX:+UseThreadPriorities -XX:ThreadPriorityPolicy=42 -Dcassandra.compaction.priority=1″

One last line would like to add.
In case when the IO is a genuine problem, you will need to add more nodes or replace disks with better performing one’s or high IO disks.

Hope It Helps
Prashant Dixit