perf(gatsby): drop severe scaling regression caused by analytics #24709
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
gatsbot
bot
added
the
status: triage needed
Regression introduced in #22851 The problem seems to be that these calls to `v8.serialize` trigger the gc to start a full hold-the-world mark-and-sweep step sooner. In a benchmark of 150k files, the step would trigger almost always after between 100k and 110k queries had run, and it would pause the process for 60+ seconds. Example benchmark results from before and after that PR: ``` info bootstrap finished - 86.758 s success Building production JavaScript and CSS bundles - 9.404s success run queries - 205.676s - 150002/150002 729.31/s success Building static HTML for pages - 142.800s - 150002/150002 1050.44/s info Done building in 451.33 sec ``` ``` info bootstrap finished - 85.933 s success Building production JavaScript and CSS bundles - 8.335s success run queries - 84.795s - 150002/150002 1769.00/s success Building static HTML for pages - 141.000s - 150002/150002 1063.84/s info Done building in 320.158 sec ``` This is very consistent behavior. We looked at the change and agreed that the best was to just drop this measurement since it was for the sake of analytics and a non-vital metric to record. We'd rather have the perf than the metric. Numbers for the fix, same benchmark, first on current master and then on this PR: ``` info bootstrap finished - 79.788s success Building production JavaScript and CSS bundles - 9.635s success run queries - 201.542s - 150002/150002 744.27/s success Building static HTML for pages - 141.535s - 150002/150002 1059.82/s info Done building in 440.766 sec ``` ``` info bootstrap finished - 80.751s success Building production JavaScript and CSS bundles - 9.570s success run queries - 87.162s - 150002/150002 1720.95/s success Building static HTML for pages - 142.609s - 150002/150002 1051.84/s info Done building in 319.151 sec ``` nice!
pvdz
force-pushed
the
drop-uniqueOperations
branch
from
eeb1826 to
7c8ce64
Compare
pvdz
added
topic: GraphQL
wardpeet
approved these changes
Jun 2, 2020
|
While this is already a good fix, I think I need to dig a little deeper. According to old benchmark dumps this is what master on February 12th did for this benchmark: So it seems we could still pump 50% faster. Hopefully that regression is also fairly easy to find. |
|
Went to repo tag And repo at tag Hard to tell what the regression might have been or how I was able to get the perf mentioned above. Could be a plugin that wasn't bumped or ... I'm not sure. But I'm pretty sure it's not worth chasing this kinda of delta over such a long timespan (over 3 months) so let's look forward. |
axe312ger
pushed a commit
that referenced
this pull request
Jun 23, 2020
) * perf(gatsby): drop severe scaling regression caused by analytics Regression introduced in #22851 The problem seems to be that these calls to `v8.serialize` trigger the gc to start a full hold-the-world mark-and-sweep step sooner. In a benchmark of 150k files, the step would trigger almost always after between 100k and 110k queries had run, and it would pause the process for 60+ seconds. Example benchmark results from before and after that PR: ``` info bootstrap finished - 86.758 s success Building production JavaScript and CSS bundles - 9.404s success run queries - 205.676s - 150002/150002 729.31/s success Building static HTML for pages - 142.800s - 150002/150002 1050.44/s info Done building in 451.33 sec ``` ``` info bootstrap finished - 85.933 s success Building production JavaScript and CSS bundles - 8.335s success run queries - 84.795s - 150002/150002 1769.00/s success Building static HTML for pages - 141.000s - 150002/150002 1063.84/s info Done building in 320.158 sec ``` This is very consistent behavior. We looked at the change and agreed that the best was to just drop this measurement since it was for the sake of analytics and a non-vital metric to record. We'd rather have the perf than the metric. Numbers for the fix, same benchmark, first on current master and then on this PR: ``` info bootstrap finished - 79.788s success Building production JavaScript and CSS bundles - 9.635s success run queries - 201.542s - 150002/150002 744.27/s success Building static HTML for pages - 141.535s - 150002/150002 1059.82/s info Done building in 440.766 sec ``` ``` info bootstrap finished - 80.751s success Building production JavaScript and CSS bundles - 9.570s success run queries - 87.162s - 150002/150002 1720.95/s success Building static HTML for pages - 142.609s - 150002/150002 1051.84/s info Done building in 319.151 sec ``` nice! * And drop the import
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.
Regression introduced in #22851
The problem seems to be that these calls to
v8.serializetrigger the gc to start a full hold-the-world mark-and-sweep step sooner. In a benchmark of 150k pages, the step would trigger almost always after between 100k and 110k queries had run, and it would pause the process for 60+ seconds.Example benchmark results from before and after that PR:
This is very consistent behavior. We looked at the change and agreed that the best was to just drop this measurement since it was for the sake of analytics and a non-vital metric to record. We'd rather have the perf than the metric.
Numbers for the fix, same benchmark, first on current master and then on this PR:
nice!