vttablet: Hot Row Protection: Make the transaction concurrency for a hot row configurable.#3123
Conversation
row configurable. By default, we will allow up to 5 transactions through to the tx pool to have a better pipelining effect and ensure that MySQL is always busy.
transaction support.
concurrent transactions from pull to push. Instead of pre-filling the channel with tokens, we will occupy one slot in the channel only while our transaction is pending. This way, we avoid the costly prepopulation of the channel with tokens.
…is a second transaction and the row actually becomes hot. This saves 22% execution time for the default case that there is no hot row. before: $ go test -bench BenchmarkTxSerializer_NoHotRow -benchtime 30s goos: darwin goarch: amd64 BenchmarkTxSerializer_NoHotRow-4 100000000 330 ns/op after: $ go test -bench BenchmarkTxSerializer_NoHotRow -benchtime 30s goos: darwin goarch: amd64 BenchmarkTxSerializer_NoHotRow-4 200000000 257 ns/op memstats before: 64 B/op 2 allocs/op memstats after: 160 B/op 3 allocs/op Note that this optimization is important for two reasons: - The code is on the critical serving path i.e. every single vttablet request has to go through it. - This code path has to use a global mutex which effectively serializes all requests and reduces the parallelism. Given that, I think this optimization is justified.
michael-berlin
force-pushed
the
hotrow_concurrency
branch
from
9664bdd to
c2c7a4b
Compare
sougou
left a comment
sougou
left a comment
There was a problem hiding this comment.
I would have ideally liked to see benchmarks that this actually makes a difference, but I'll trust your instincts on this one :).
|
I did some internal benchmarks on this: In case of 100 concurrent transactions hitting the same row, I saw a 4% performance regression which went away with this fix. In general, this fix makes sense: If we allow only a concurrency of 1, all transactions are stuck at a mutex in the vttablet code and not a MySQL mutex. I.e. when we unblock the transaction in vttablet, it takes time until it goes through the stack and is ready in MySQL. By having several transactions always ready in MySQL (I randomly picked a default of 5 here), we avoid this problem and have a pipelining effect. |
|
I'm merging this now. unit_race keeps failing due to unrelated, flaky errors. |
3 participants
By default, we will allow up to 5 transactions through to the tx pool to have a better pipelining effect and ensure that MySQL is always busy.
Sugu, I recommend to review this change one commit at a time. Some guidance: