TWCS experiments and improvement proposals

This JIRA item describes experiments with DateTieredCompactionStartegy (DTCS) and TimeWindowCompactionStrategy (TWCS) and proposes modifications to the TWCS. In a test system several crashes were caused intentionally (and unintentionally) and repair operations were executed leading to flood of small SSTables. Target was to be able compact those files are release disk space reserved by duplicate data. Setup is following:

• Three nodes
• DateTieredCompactionStrategy, max_sstable_age_days = 5
  Cassandra 2.1.2

The setup and data format has been documented in detailed here https://issues.apache.org/jira/browse/CASSANDRA-9644.

The test was started by dumping few days worth of data to the database for 100 000 signals. Time graphs of SStables from different nodes indicates that the DTCS has been working as expected and SStables are nicely ordered in time wise.

See files:
node0_20150727_1250_time_graph.txt
node1_20150727_1250_time_graph.txt
node2_20150727_1250_time_graph.txt

Status=Up/Down

All nodes crashed due to power failure (know beforehand) and repair operations were started for each node one at the time. Below is the behavior of SSTable count on different nodes. New data was dumped simultaneously with repair operation.

SEE FIGURE: sstable_count_figure1.png

Vertical lines indicate following events.

1) Cluster was down due to power shutdown and was restarted. At the first vertical line the repair operation (nodetool repair -pr) was started for the first node
2) Repair for the second repair operation was started after the first node was successfully repaired.
3) Repair for the third repair operation was started
4) Third repair operation was finished
5) One of the nodes crashed (unknown reason in OS level)
6) Repair operation (nodetool repair -pr) was started for the first node
7) Repair operation for the second node was started
8) Repair operation for the third node was started
9) Repair operations finished

These repair operations are leading to huge amount of small SSTables covering the whole time span of the data. The compaction horizon of DTCS was limited to 5 days (max_sstable_age_days) due to the size of the SStables on the disc. Therefore, small SStables won't be compacted. Below are the time graphs from SSTables after the second round of repairs.

Status=Up/Down

See files:
node0_20150810_1017_time_graph.txt
node1_20150810_1017_time_graph.txt
node2_20150810_1017_time_graph.txt

To get rid of the SStables the TimeWindowCompactionStrategy was taken into use. Window size was set to 5 days. Cassandra version was updated to 2.1.8. Below figure shows the behavior of SStable count. TWCS was taken into use 10.8.2015 at 13:10. The maximum amount of files to be compacted in one task was limited to 32 files to avoid running out of disk space.

See Figure: sstable_count_figure2.png

Shape of the trend indicates clearly how selection of SStables for buckets based on size affects. Combining files gets slower when files are getting bigger inside the time window. When the time window does not have any more compactions to be done the next time window is started. Combining small files is again fast and the number of SStables decreases quickly. Below are the time graphs for SStables when compactions were ready with TWCS. New data was not dumped simultaneously with compactions.

See files:
node0_20150812_1531_time_graph.txt
node1_20150812_1531_time_graph.txt
node2_20150812_1531_time_graph.txt

Datacenter: datacenter1
=======================
Status=Up/Down

Data dumping was activated again and the SStable statistics were observed again on the next morning.

See files:
node0_20150813_0835_time_graph.txt
node1_20150813_0835_time_graph.txt
node2_20150813_0835_time_graph.txt

Since the data was dumped to the history the newest data did not come into the current time window that is determined from the system time. Since new small SSTables (approximately 30- 50 MB in size) are appearing continuously the compaction ended up compacting together one large SStable with several small files. The code was modified so that the current time is determined from the newest time stamp in the SStables (like in DTCS).This modification led to much more reasonable compaction behavior for the case when historical data is pushed to the database. Below are the time grahps from nodes after one day. Size tiered compaction was now able to work with newest files as intended while dumping data in real-time.

See files:
node0_20150814_1054_time_graph.txt
node1_20150814_1054_time_graph.txt
node2_20150814_1054_time_graph.txt

The change in behavior is clearly visible in the compaction hierarchy graph below. TWCS modification is visible starting from the line 39. See the description of the file format in https://issues.apache.org/jira/browse/CASSANDRA-9644.

See file: 20150814_1027_compaction_hierarchy.txt

The behavior of the TWCS looks really promising and works also in practice!!!
We would like to propose some ideas for future development of the algorithm.

1) The current time window would be determined from the newest time stamp found in SSTables. This allows the effective compaction of the SSTables when data is fed to the history in timely order. In dumping process the time stamp of the column is set according to the time stamp of the data sample.

2) The count of SSTables participating in one compaction could be limited either by the number of files given by max_threshold OR by the sum of size of files selected for the compaction bucket. File size limitation would prevent combining a large files together potentially causing out of disk space situation or extremely long lasting compaction tasks.

3) Now time windows are handled one by one starting from the newest. This will not lead to the fastest decrease in SStable count. An alternative might a round-robin approach in which time windows are stepped through and only one compaction task for that given time window is done and then moving to the next time window.

Side note: while observing the compaction process it appears that compaction is intermittently using two threads for the compaction. However, sometimes during a long lasting compaction task (hours) another thread was not kicking in and working with small SSTables even if there were thousands of those available for compactions.