Uploaded image for project: 'HBase'
  1. HBase
  2. HBASE-6093

Flatten timestamps during flush and compaction


    • Type: New Feature
    • Status: Resolved
    • Priority: Minor
    • Resolution: Incomplete
    • Affects Version/s: None
    • Fix Version/s: None
    • Component/s: io, Performance, regionserver
    • Labels:


      Many applications run with maxVersions=1 and do not care about timestamps, or they will specify one timestamp per row as a normal KeyValue rather than per-cell.

      Then, DataBlockEncoders like those in HBASE-4218 and HBASE-4676 often encode timestamps as diffs from the previous or diffs from the minimum timestamp in the block. If all timestamps in a block are the same, they will all compress to basically <= 8 bytes total per block. This can be 10% to 25% space savings for some schemas, and that savings is realized both on disk and in block cache.

      We could add a ColumnFamily setting flattenTimestamps=[true/false]. If true, then all timestamps are modified during a flush/compaction to the currentTimeMillis() at the start of the flush/compaction. If all timestamps are made identical in a file, then the encoder will be able to eliminate them.

      The simplest use case is probably that where all inserts are type=Put, there are no overwrites, and there are no deletes. As use cases get more complex, then so does the implementation.

      For example, what happens when there is a Put and a Delete of the same cell in the same memstore? Maybe for a flush at t=flushStartTime, the Put gets timestamp=t, and the Delete gets timestamp=t+1. Or maybe HBASE-4241 could take care of this problem.




            • Assignee:
              mcorgan Matt Corgan
            • Votes:
              2 Vote for this issue
              7 Start watching this issue


              • Created: