wal: use page buffered writer for writing records

[heyitsanthony]

Fixes #6271

/cc @gyuho

[gyuho]

Ok I will try this branch in failpoints cluster.

[xiang90]

define 4096 somewhere?

[gyuho]

Also doc that this is buffering same as bufio https://github.com/golang/go/blob/master/src/bufio/bufio.go#L18?

[heyitsanthony]

@gyuho it's only a coincidence the wal page size is the bufio default buffer size; the full buffer size for a pagewriter is 128k. I'll add a comment about the choice of the value.

[gyuho]

Got it. Thanks!

[xiang90]

LGTM. Have you tried to run the benchmark in wal pkg to see if there is any performance gain?

[heyitsanthony]

@xiang90 benchmark found a bug in the page writer 👍

Master:

BenchmarkWrite100EntryWithoutBatch-4             200      10144426 ns/op       0.01 MB/s
BenchmarkWrite100EntryBatch10-4                 2000        943462 ns/op       0.11 MB/s
BenchmarkWrite100EntryBatch100-4               10000        116260 ns/op       0.93 MB/s
BenchmarkWrite100EntryBatch500-4               50000         28449 ns/op       3.80 MB/s
BenchmarkWrite100EntryBatch1000-4             100000         15095 ns/op       7.15 MB/s
BenchmarkWrite1000EntryWithoutBatch-4            100      10645949 ns/op       0.09 MB/s
BenchmarkWrite1000EntryBatch10-4                2000       1016150 ns/op       0.99 MB/s
BenchmarkWrite1000EntryBatch100-4              10000        144891 ns/op       6.96 MB/s
BenchmarkWrite1000EntryBatch500-4              50000         42404 ns/op      23.79 MB/s
BenchmarkWrite1000EntryBatch1000-4             50000         28502 ns/op      35.40 MB/s

This patch:

BenchmarkWrite100EntryWithoutBatch-4             100      10629690 ns/op       0.01 MB/s
BenchmarkWrite100EntryBatch10-4                 2000        963754 ns/op       0.11 MB/s
BenchmarkWrite100EntryBatch100-4               10000        107374 ns/op       1.01 MB/s
BenchmarkWrite100EntryBatch500-4              100000         23026 ns/op       4.69 MB/s
BenchmarkWrite100EntryBatch1000-4             200000         11675 ns/op       9.25 MB/s
BenchmarkWrite1000EntryWithoutBatch-4            100      10485759 ns/op       0.10 MB/s
BenchmarkWrite1000EntryBatch10-4                2000        997684 ns/op       1.01 MB/s
BenchmarkWrite1000EntryBatch100-4              10000        121113 ns/op       8.33 MB/s
BenchmarkWrite1000EntryBatch500-4              50000         25324 ns/op      39.84 MB/s
BenchmarkWrite1000EntryBatch1000-4            100000         13901 ns/op      72.58 MB/s

[gyuho]

I can confirm that this fixes #6271 since it used to fail with less than 3 rounds. Now it's over 10 rounds. Thanks!

[xiang90]

lgtm