Cache Writer optimization

We want to optimize the cache writer to support:

• parallel cache writes for tap feed entries

Non-sparse Cache Design

####Data Model Cache Block

Key format: _cache_block:[channelName]:[blockIndex]

Raw storage, with the format

• Sequence offset- 8 bytes. Starting sequence for the block.
• Sequence array - 1 byte per sequence, indexed as (sequence - sequenceOffset). Currently only using two bits of each byte for information:
  • bit 0 - flag for whether the sequence is in this channel
  • bit 1 - flag for whether the sequence indicates a removal from the channel

Cache clock

Key = _cache_clock:[channelname]

Integer value - incremented whenever a sequence is added to the channel

Cache entry

Key: _cache_entry:[sequence] JSON document containing LogEntry information for a sequence returned by the _changes feed: DocID, RevID, Sequence, Flags

####Write For each tap feed entry:

• 1 write for the _cache:entry document
  • stores the information sent by the _changes feed (DocID, RevID, Sequence, Flags)
• per channel:
  • 1 read for the cache block
  • 1 CAS write for the cache block
  • 1 incr for the cache clock

Total ops per feed entry: 1 + (# channels * 3)

####Read For each poll iteration:

• per channel:
  • 1 read of the cache clock
  • if clock shows changes
    • 1 read of cache block (low potential for additional cache block read):
      • per sequence:
        • read of _cache:entry document

Total ops per iteration: (# channels) * (2 + (#changes))

Benefits:

• automatic sequence ordering in cache blocks
• identify target block for sequence by key - no additional call to look up which block a sequence is in

Cons:

• Read/CAS write needed for update
  • for a single streamed writer, it's slow
  • for multi-streamed or multiple writers, high contention on busy channels
• unused storage for sparse channels, could be as low as 1 file per sequence