Resiliency in Elasticsearch and Lucene

As Elasticsearch clusters grow larger, their resilience to hardware and network failure becomes increasingly important. At Elasticsearch, we invest a LOT in making both Elasticsearch and Apache Lucene both detect and cope with increasingly complex failures.

In this talk we will go through recent highlights and some of the future directions we plan to follow. We will touch all aspects of Elasticsearch, ranging from the lowest level of a single file, through network connection of a single node, and all the way up distributed failures on the cluster level. Even though the talk is about possible failures and various coping strategies, you will also learn about the inner workings of Elasticsearch - an interesting peek under the hood.

Talk given at Elastic{ON} 2015, by Igor Motov & Boaz Leskes

Transcript

1. Resiliency in Elasticsearch and Lucene Boaz Leskes and Igor Motov
2. { } CC-BY-ND 4.0 Resiliency noun 1. the power or ability to return to the original form, position, etc., after being bent, compressed, or stretched; elasticity. 2. ability to recover readily from illness, depression, adversity, or the like; buoyancy. 2 re·sil·ience
3. { } CC-BY-ND 4.0 Failures happen 3
4. { } CC-BY-ND 4.0 Why is it so important now? 4
5. { } CC-BY-ND 4.0 Average Elasticsearch cluster growth 5 Large cluster circa 2011
6. { } CC-BY-ND 4.0 Average Elasticsearch cluster growth 6 Large cluster circa 2011 Large cluster circa 2015
7. { } CC-BY-ND 4.0 Failure rates X number of nodes 7
8. { } CC-BY-ND 4.0 8 Slow Fast Node Cluster kill -9 dead disk corruption long GC master gone network disconnects timeouts
9. { } CC-BY-ND 4.0 9 Slow Fast Node Cluster kill -9 dead disk corruption long GC master gone network disconnects timeouts SOFTWARE BUGS
10. { } CC-BY-ND 4.0 10 Slow Fast Node Cluster kill -9 dead disk corruption long GC master gone network disconnects timeouts
11. { } CC-BY-ND 4.0 11 Work in progress
12. { } CC-BY-ND 4.0 Pulling the plug 12
13. { } CC-BY-ND 4.0 Replicas and Transaction Log 13 Replica Shard Lucene Index Transaction Log Lucene Buffer Primary Shard Lucene Index Transaction Log Lucene Buffer
14. { } CC-BY-ND 4.0 Replicas and Transaction Log 14 Replica Shard Lucene Index Transaction Log Lucene Buffer Primary Shard Lucene Index Transaction Log Lucene Buffer
15. { } CC-BY-ND 4.0 Replicas and Transaction Log 15 Replica Shard Lucene Index Transaction Log Lucene Buffer Primary Shard Lucene Index Transaction Log Lucene Buffer
16. { } CC-BY-ND 4.0 Replicas and Transaction Log 16 Replica Shard Lucene Index Transaction Log Lucene Buffer Primary Shard Lucene Index Transaction Log Lucene Buffer
17. { } CC-BY-ND 4.0 Transaction Log • Transaction Log – stores every operation (create/update/delete) – fsync-ed every 5 sec (configurable) • Lucene Segments – fsync-ed when transaction log is full (every 30 min or 200mb) 17
18. { } CC-BY-ND 4.0 Hard Disk Failures 18
19. { } CC-BY-ND 4.0 Hard Drive Failures • Complete failure • Running out of disk space • Data corruption 19
20. { } CC-BY-ND 4.0 Complete Disk Failures • Automatic shard failover • Replicas 20
21. { } CC-BY-ND 4.0 Multi data paths 21 Disk 1 Disk 2 Disk3 Disk 4 shard 1 shard 2 shard 3 shard 4
22. { } CC-BY-ND 4.0 Multi data paths 22 Disk 1 Disk 2 Disk3 Disk 4 shard 1 shard 2 shard 3 shard 4
23. { } CC-BY-ND 4.0 Multi data paths 23 Disk 1 Disk 2 Disk3 Disk 4 shard 1 shard 2 shard 3 shard 4
24. { } CC-BY-ND 4.0 Complete Disk Failures (WIP) • Current multi-path setup strips shard data across multiple disks • Disk loss impacts all shards on node • Reduce failure impact by using one disk per shard 
 (v2.0, #9498) 24
25. { } CC-BY-ND 4.0 Running out of Disk Space • Can lead to truncated files and thus corruption • Easy to anticipate by monitoring • Disk-space aware allocation decider – added in 0.90.4 – enabled by default since 1.3.5 25
26. { } CC-BY-ND 4.0 file content (10GB) Data Corruption - the bit in the haystack 26
27. { } CC-BY-ND 4.0 Data Corruption - Checksums 27 footer + checksum file content (10gb)
28. { } CC-BY-ND 4.0 Data Corruption • Elasticsearch automatically checks checksums of – Small files on index open (since v1.3.0) – All files during replication, relocation, snapshot, and restore (since v1.3.3 - v1.4.0) – Transaction log (since v1.4.0) – Use checksum to identify entire segments to reduce chance of hash collisions (since v1.4.0) 28
29. { } CC-BY-ND 4.0 Data Corruption (WIP) • Checksums of Metadata files (coming in v1.5.0) • Support validation of checksum on all files when node starts (v2.0.0, #9183) • Make validation during merge operation more efficient (v2.0.0, LUCENE-5894) • Add per-segment/per-commit ids (v2.0.0, LUCENE-5895) • Prevent use of known-bad java versions (v1.5.0, #7580) 29
30. { } CC-BY-ND 4.0 Cluster Issues 30
31. { } CC-BY-ND 4.0 Why nodes leave the cluster? • Complete node failure • Unresponsive nodes • Network Failures 31
32. { } CC-BY-ND 4.0 Unresponsive Nodes 32
33. { } CC-BY-ND 4.0 Biggest Memory User- Field Data • sorting • aggregations • doc[“foo”] in scripts • Parent-child id cache – has_child/has_parent queries 33
34. { } CC-BY-ND 4.0 Circuit Breakers • Estimate size of the field data for each query and fail the query if it tries to load too much data – field data (since v1.0.0) – parent-child (since v1.1.0) – some aggregation structures (since v1.4.0) 34
35. { } CC-BY-ND 4.0 Doc values • On-disk low memory alternative to field data • Significant performance improvements in v1.4.0 35
36. { } CC-BY-ND 4.0 OOM Resiliency (WIP) • Add hard limit on from/size (v1.5.0, #9311) • Add hit size circuit breaker (TBD, #9310) • Prevent combinatorial explosion in aggregations (TBD, #8081) • Smarter filter caching (LUCENE-6303, ES 2.0) 36
37. { } CC-BY-ND 4.0 Dedicate Master Nodes 37 node 1 node.master: false node.data: true node 1 node 1 node 1 node 1 node 1 node 1 node 1 node 1 node 1 node 1 data 1 node.master: true node.data: false master 1 master 2 master 3
38. { } CC-BY-ND 4.0 Network Issues 38
39. { } CC-BY-ND 4.0 Partitions & partial knowledge 39 node M node node node node
40. { } CC-BY-ND 4.0 Remember to set minimum_master_nodes • Set discovery.zen.minumum_master_nodes:(N/2 +1) in elasticsearch.yml 40
41. { } CC-BY-ND 4.0 Partitions & partial knowledge 41 node M node node node node
42. { } CC-BY-ND 4.0 Improved Zen Discovery • Significant improvements in v1.4.0 – Gossip on master loss – bigger ping outreach – resiliency to stale gossip – better two-masters resolution – faster failure detection 42
43. { } CC-BY-ND 4.0 Improving Zen Discovery (WIP) • Prevent setting incorrect minimum_master_nodes (v1.5.0, #8321, #9051) • Refuse revived master (v1.5.0) 43
44. { } CC-BY-ND 4.0 External pressure 44
45. { } CC-BY-ND 4.0 External pressure - bounded queues and thread pools (<0.20) - time out long running queries (WIP, PR #9156) - index throttling (v1.2.0, #6066) 45
46. { } CC-BY-ND 4.0 Known Unknowns 46
47. { } CC-BY-ND 4.0 Known Unknowns - Simulate disruptions (v1.4.0, #7492) - Simulate corruption (v1.3.0, #5924) - Reproducible evil - Users info is critical 47
48. { } CC-BY-ND 4.0 It’s ongoing effort • Check the progress on our resiliency status page – http://www.elasticsearch.org/guide/en/elasticsearch/ resiliency/current/ • Or search for issues labeled “resiliency” on github – https://github.com/elasticsearch/elasticsearch/ 48
49. { } Thank you! boaz@elastic.co, igor@elastic.co @bleskes, @imotov
50. { } This work is licensed under the Creative Commons Attribution-NoDerivatives 4.0 International License. To view a copy of this license, visit: http://creativecommons.org/licenses/by-nd/4.0/ or send a letter to: Creative Commons PO Box 1866 Mountain View, CA 94042 USA CC-BY-ND 4.0