Support reading profiles from JsonSlabs files#6037
Conversation
We'll want to add a binary serialization soon, but we'll also keep the JSON serializatios as an option. This rename makes it clear which one you're calling.
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #6037 +/- ##
==========================================
- Coverage 83.78% 83.77% -0.01%
==========================================
Files 329 329
Lines 34544 34557 +13
Branches 9675 9668 -7
==========================================
+ Hits 28941 28950 +9
- Misses 5174 5178 +4
Partials 429 429 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
Fixes firefox-devtools#6024. JsonSlabs is an alternative serialization format that's more efficient than JSON if the serialized object contains typed arrays. See https://github.com/mstange/json-slabs/ for details. Our profiles currently do not yet contain typed arrays. But with this in place, we can start converting more and more tables / columns to use typed arrays, and incrementally reap the efficiency benefits. This patch only adds the reading. In the UI, uploading / downloading profiles still uses JSON. For the profiler-edit node script, this patch also adds the "writing": If you run profiler-edit with `-o some-filename-ending-in.jslb`, then it will create a JSLB file (or a compressed JSLB file for .jslb.gz). --- The profile format "structure" is still the same as before. The format version is still given by profile.meta.preprocessedProfileVersion. Loading a JSLB file with a profile.meta.preprocessedProfileVersion that's higher than the currently known one will still complain, and attempt to reload the page or the service worker. This reload behavior is the primary reason why I want to get the "reading" part taken care of first, before we start converting any parts of the profile to get efficiency wins: The sooner we land this, the more likely it will be that, in the future, if somebody has a newer JSLB file but an old cached profiler, they'll see a useful error and get the auto-reload. Without this patch, loading a JSLB file in the profiler gives the following error: > Error: Unserializing the profile failed: Error: The profile array > buffer could not be parsed as a UTF-8 string. --- With all that said, it turns out there is already a case where this format gives an efficiency win, thanks to its "split out" feature that lets us put certain subtrees of the profile into separate JSON slabs: If the original JSON is larger 512MiB, we can split it into individual slabs which are each smaller than 512MiB, so the built-in JSON.parse will work and we don't have to use the slow streaming parse. Example: https://storage.googleapis.com/profiler-get-symbols-fixtures/large-speedometer3-profile.json.gz expands to a 577.74MB JSON file, https://storage.googleapis.com/profiler-get-symbols-fixtures/large-speedometer3-profile.jslb.gz expands to a 564.66MB JSLB file (slightly smaller because compacting removed unused strings) The JSLB file breaks down as follows: ``` idx type bytes elements path --- ------- ---------- -------- ------------------- 0 json 16.76 MiB - . (root) 1 json 264.82 MiB - .shared.stackTable 2 json 91.01 MiB - .shared.frameTable 3 json 8.47 MiB - .shared.funcTable 4 json 28.80 MiB - .shared.stringArray 5 json 154.80 MiB - .threads --- ------- ---------- -------- ------------------- 6 slabs 564.66 MiB ``` All of those individual JSON slabs are now under 512MiB. Loading the JSON file, uses streaming parser: https://share.firefox.dev/3PcB9vK (19 seconds) Loading the JSLB file, uses native JSON.parse: https://share.firefox.dev/3RkFcXs (2.3 seconds, 8.3x faster)
mstange
force-pushed
the
push-pxlosymwykls
branch
from
d7175f8 to
206f8ff
Compare
canova
left a comment
canova
left a comment
There was a problem hiding this comment.
Thanks for this work, it looks like great change toward better file format.
This patch only adds the reading. In the UI, uploading / downloading profiles still uses JSON.
Ah, that's a good idea. I initially thought this PR was changing the defaults too. But now it looks a lot less scary :)
I read the json-slabs package code as well and reviewed it and it looks good to me as well. Thanks for implementing these and improving our file format!
With all that said, it turns out there is already a case where this format gives an efficiency win, thanks to its "split out" feature that lets us put certain subtrees of the profile into separate JSON slabs: If the original JSON is larger 512MiB, we can split it into individual slabs which are each smaller than 512MiB, so the built-in JSON.parse will work and we don't have to use the slow streaming parse.
Oh, that's a pretty good win!
|
Note that this PR by itself was only supporting jslb files in the "from-file" path, i.e. if you drag a file into the profiler or enter it via the file input. It wasn't working when downloading profiles from profile storage (for "public") or when passing a URL via "from-url". #6072 fixes that. |
2 participants
Deploy preview (load this example jslb in it)
Fixes #6024.
JsonSlabs is an alternative serialization format that's more efficient than JSON if the serialized object contains typed arrays.
See https://github.com/mstange/json-slabs/ for details.
Our profiles currently do not yet contain typed arrays. But with this in place, we can start converting more and more tables / columns to use typed arrays, and incrementally reap the efficiency benefits.
This patch only adds the reading. In the UI, uploading / downloading profiles still uses JSON.
For the profiler-edit node script, this patch also adds the "writing": If you run profiler-edit with
-o some-filename-ending-in.jslb, then it will create a JSLB file (or a compressed JSLB file for .jslb.gz).The profile format "structure" is still the same as before. The format version is still given by profile.meta.preprocessedProfileVersion. Loading a JSLB file with a profile.meta.preprocessedProfileVersion that's
higher than the currently known one will still complain, and attempt to reload the page or the service worker.
This reload behavior is the primary reason why I want to get the "reading" part taken care of first, before we start converting any parts of the profile to get efficiency wins: The sooner we land this, the more likely it will be that, in the future, if somebody has a newer JSLB file but an old cached profiler, they'll see a useful error and
get the auto-reload.
Without this patch, loading a JSLB file in the profiler gives the following error:
With all that said, it turns out there is already a case where this format gives an efficiency win, thanks to its "split out" feature that lets us put certain subtrees of the profile into separate JSON slabs: If the original JSON is larger 512MiB, we can split it into individual slabs which are each smaller than 512MiB, so the built-in JSON.parse will work and we don't have to use the slow streaming parse.
Example:
https://storage.googleapis.com/profiler-get-symbols-fixtures/large-speedometer3-profile.json.gz
expands to a 577.74MiB JSON file,
https://storage.googleapis.com/profiler-get-symbols-fixtures/large-speedometer3-profile.jslb.gz
expands to a 564.66MiB JSLB file (slightly smaller because compacting removed unused strings)
The JSLB file breaks down as follows:
All of those individual JSON slabs are now under 512MiB.
Loading the JSON file, uses streaming parser: https://share.firefox.dev/3PcB9vK (19 seconds)
Loading the JSLB file, uses native JSON.parse: https://share.firefox.dev/3RkFcXs (2.3 seconds, 8.3x faster)
The example jslb next to the deploy preview link was generated as follows:
yarn build-node-tools && node ./node-tools-dist/profiler-edit.js --from-hash 29apty4565r8dec0hrjfty68w88rexc6na6bn1r -o ~/Downloads/normal-profile.jslb.gzFollow-ups: