async-flumelog

Async flumelog

This module is heavily inspired by flumelog-aligned-offset. It is an attempt to implement the same concept but in a simpler fashion, making it easier to reason about the code. Flumelog is the lowest part of the SSB stack, so it should extremly stable while still maintaining good performance.

An async flumelog consists of a number of blocks, that contain a number of records. A record is simply it's length, as a 16-bit unsigned integer, followed by the data bytes. A record must be in one and only one block, which means there probably will be some empty space at the end of a block. Blocks are always written in full.

<block>
  <record
    <length: UInt16LE>
    <data: Bytes>
  </record>*
</block>*

In contrast to flumelog-aligned-offset there is no additional length after the data in a record and no pointer at the end of a block. These were there to be able to iterate over the log in reverse, but I have never seen the need for that.

Writing to the log is always async. Note this is different from flumelog-offset and flumelog-aligned-offset. The since observable will be updated once the data is written. The onDrain callback can be used to know when data has been written if needed. Streaming will only emit values that have been written to storage. This is to ensure that a view will never get ahead of the main log and thus end up in a bad state if the system crashes before data is written. get will return values that have not been written to disk yet.

Benchmarks

Running bench-flumelog reveals the following numbers. Async flumelog is faster that regular flumelog-offset in all categories. The most important numbers are append (used for onboarding) and stream (used for building indexes). Flumelog-aligned-offset is not included in the benchmarks, as it writes every message synchronously rendering the results invalid.

async flumelog:

name, ops/second, mb/second, ops, total-mb, seconds
append, 923964.807, 138.002, 4620748, 690.149, 5.001
stream, 1059075.865, 158.182, 4620748, 690.149, 4.363
stream no cache, 1102803.818, 164.713, 4620748, 690.149, 4.19
stream10, 2540947.641, 379.51, 12714902, 1899.068, 5.004
random, 39715.656, 5.931, 198618, 29.664, 5.001

flumelog offset:

name, ops/second, mb/second, ops, total-mb, seconds
append, 306180.037, 45.74, 3064556, 457.817, 10.009
stream, 294511.348, 43.997, 2945408, 440.017, 10.001
stream no cache, 327724.949, 48.959, 3064556, 457.817, 9.351
stream10, 452973.302, 67.67, 4530186, 676.776, 10.001
random, 28774.712, 4.298, 287891, 43.008, 10.005

To run the benchmarks the small bench-flumelog.patch needs to be applied.

JITDB results for more real world benchmarks are available as jitdb-results.