Skip to content

performance.md

Latest commit

 

History

History
50 lines (49 loc) · 2.31 KB

performance.md

File metadata and controls

50 lines (49 loc) · 2.31 KB

#Performance

Performance considerations

  • Right compaction strategy?
    • SizeTieredCompactionStrategy
      • Compaction happens when SSTable reaches a certain size
      • Not good for frequently updated cell in a (wide) row - multiple seeks to get entire row
      • Good for write-once data
      • Requires at least as much free disk space for compaction as the size of the largest column family
    • LeveledCompactionStrategy
      • Predictable read latency for rows which are frequently updated and deleted
      • SSTables of a fixed size grouped into levels
      • Each level 10x size of previous level (eg. L0 is ~5 MB, L1 is ~50 MB)
      • Within each level, SSTables are guaranteed to be non-overlapping
    • DateTieredCompactionStrategy
      • All rows in a certain time range stored in a single SSTable
      • All values set to expire at a certain time stored in a single SSTable
  • Right compression mode?
    • No compression may mean less CPU usage, but more storage, more network if using network store like EBS
  • Right consistency level?
    • You can give up consistency for faster response and availability
  • Right data model?
    • Wrong use of secondary index?
    • Using ALLOW FILTERING?
  • Reading huge volumes of data in a single query?
    • Buffers may take up memory in Cassandra and in the clients
    • Network timeout may trigger a retry and magnify the issue
  • Right data access pattern?
    • Using Cassandra as queue is slow because "Delete" just creates a tombstone. Range queries read them and skip the values.
    • Read before write
  • Commit log and data are in same disk?
  • nodetool repair being run regularly?
  • Using vnodes?
    • Eases load when nodes are added and removed

Performance tuning and monitoring

  • Reference
  • Enable tracing
  • nodetool
    • cfhistograms
    • cfstats
    • tpstats
  • OS
    • top
    • iotop
    • iftop
    • vmstat
    • iostat
  • cassandra-stress
    • Remember to run for at least an hour to ensure effects of compaction are taken into account.
  • Write Survey Mode