sql: lower default sampling rate to 1% #62998
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
This change is |
|
Raw results for the curious. Group 10.0 runs:0.01 runs:0.1 runs:Group 20.00 runs0.01 runs:0.1 runs: |
We arrived at the previous default rate of 10% back in cockroachdb#59379. This was back when we were creating real tracing spans for all statements, and for sampled statements, we were propagating additional stats payloads. Consequently what cockroachdb#59379 ended up measuring (and finding the overhead acceptable) for was the performance hit we would incur for propagating stats payloads for statements already using real tracing spans. Since then, the landscape has changed. Notably we introduced cockroachdb#61777, which made it so that we were only using real tracing spans for sampled statements. This was done after performance analysis in cockroachdb#59424 showed that the use of real tracing spans in all statements resulted in tremendous overhead, for no real benefit. What this now leaves us with is a sampling rate that was tuned by only considering the stats payload overhead. What we want now is to also consider the overhead of using real tracing spans for sampled statements, vs. not. Doing this analysis gives us a very different picture for what the default rate should be. --- To find out what the overhead for sampled statements are currently, we experimented with kv95/enc=false/nodes=1/cpu=32. It's a simple benchmark that does little more than one-off statements, so should give us a concise picture of the sampling overhead. We ran six experiments in total (each corresponding to a pair of read+write rows), done in groups of three (each group corresponding to a table below). Each run in turn is comprised of 10 iterations of kv95, and what's varied between each run is the default sampling rate. We pin a sampling rate of 0.0 as the baseline that effectively switches off sampling entirely (and tracing), and measure the throughput degradation as we vary the sampling rate. ops/sec ops/sec --------------------|------------------|------------------ rate op grp | median diff | mean diff --------------------|------------------|------------------ 0.00 / read / #1 | 69817.90 | 69406.37 0.01 / read / #1 | 69300.35 -0.74% | 68717.23 -0.99% 0.10 / read / #1 | 67743.35 -2.97% | 67601.81 -2.60% 0.00 / write / #1 | 3672.55 | 3653.63 0.01 / write / #1 | 3647.65 -0.68% | 3615.90 -1.03% 0.10 / write / #1 | 3567.20 -2.87% | 3558.90 -2.59% ops/sec ops/sec --------------------|------------------|------------------ rate op grp | median diff | mean diff --------------------|------------------|------------------ 0.00 / read / #2 | 69440.80 68893.24 0.01 / read / #2 | 69481.55 +0.06% 69463.13 +0.82% (probably in the noise margin) 0.10 / read / #2 | 67841.80 -2.30% 66992.55 -2.76% 0.00 / write / #2 | 3652.45 3625.24 0.01 / write / #2 | 3657.55 -0.14% 3654.34 +0.80% 0.10 / write / #2 | 3570.75 -2.24% 3526.04 -2.74% The results above suggest that the current default rate of 10% is too high, and a 1% rate is much more acceptable. --- The fact that the cost of sampling is largely dominated by tracing is extremely unfortunate. We have ideas for how that can be improved (prototyped in cockroachdb#62227), but they're much too invasive to backport to 21.1. It's unfortunate that we only discovered the overhead this late in the development cycle. It was due to two major reasons: - cockroachdb#59992 landed late in the cycle, and enabled tracing for realsies (by propagating real tracing spans across rpc boundaries). We had done sanity checking for the tracing overhead before this point, but failed to realize that cockroachdb#59992 would merit re-analysis. - The test that alerted us to the degradation (tpccbench) had be persistently failing for a myriad of other reasons, so we didn't learn until too late that tracing was the latest offendor. tpccbench also doesn't deal with VM overload well (something cockroachdb#62361 hopes to address), and after tracing was enabled for realsies, this was the dominant failure mode. This resulted in perf data not making it's way to roachperf, which further hid possible indicators we had a major regression on our hands. We also didn't have a healthy process looking at roachperf on a continual basis, something we're looking to rectify going forward. We would've picked up on this regression had we been closely monitoring the kv95 charts. Release note: None
irfansharif
force-pushed
the
210401.lower-sample-rate
branch
from
47f586c to
f4ec58f
Compare
|
Separately I also tried grabbing tpccbench numbers, but it's pretty hard to make heads or tails from it the numbers tend to be nonsensical once the VMs plunge into overloaded territory (see #62039 (comment)). |
|
So the short summary here, if I'm understanding correctly, is that: 1% sampling gives about 1% degradation over no sampling right? |
|
Yup. |
|
bors r+ |
|
Woops, wrong PR tab. bors r- |
|
Canceled. |
yuzefovich
approved these changes
Apr 2, 2021
There was a problem hiding this comment.
What's the difference between group 1 and 2? Simply two different clusters? Or two different machine types?
Reviewed 1 of 1 files at r1.
Reviewable status:complete! 1 of 0 LGTMs obtained (waiting on @tbg)
|
No real difference, just wanted to do the whole thing over again to see that I wasn't deluding myself. TFTR! bors r+ |
|
Build succeeded: |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
We arrived at the previous default rate of 10% back in #59379. This was
back when we were creating real tracing spans for all statements, and
for sampled statements, we were propagating additional stats payloads.
Consequently what #59379 ended up measuring (and finding the overhead
acceptable) for was the performance hit we would incur for propagating
stats payloads for statements already using real tracing spans.
Since then, the landscape has changed. Notably we introduced #61777,
which made it so that we were only using real tracing spans for sampled
statements. This was done after performance analysis in #59424 showed
that the use of real tracing spans in all statements resulted in
tremendous overhead, for no real benefit.
What this now leaves us with is a sampling rate that was tuned by only
considering the stats payload overhead. What we want now is to also
consider the overhead of using real tracing spans for sampled
statements, vs. not. Doing this analysis gives us a very different
picture for what the default rate should be.
To find out what the overhead for sampled statements are currently, we
experimented with kv95/enc=false/nodes=1/cpu=32. It's a simple
benchmark that does little more than one-off statements, so should give
us a concise picture of the sampling overhead. We ran six experiments
in total (each corresponding to a pair of read+write rows), done in
groups of three (each group corresponding to a table below). Each
run in turn is comprised of 10 iterations of kv95, and what's varied
between each run is the default sampling rate. We pin a sampling rate of
0.0 as the baseline that effectively switches off sampling entirely (and
tracing), and measure the throughput degradation as we vary the sampling
rate.
The results above suggest that the current default rate of 10% is too
high, and a 1% rate is much more acceptable.
The fact that the cost of sampling is largely dominated by tracing is
extremely unfortunate. We have ideas for how that can be improved
(prototyped in #62227), but they're much too invasive to backport to
21.1.
Release note: None