i see removeEntry* and schema upgrade stuff using it in http://mxr.mozilla.org/mozilla-central/source/toolkit/components/satchel/nsFormHistory.js and schema upgrade stuff in .jsm
Marco/David, should code switching to the new async-only storage interface to prevent these kinds of omissions?
Async-only storage interface will make sure that you don't have sync code mixed with async (except for closing, see bug 874814), so I guess this would help. Mak knows that codebase better than I do, of course.
There are various reasons that disallow us from doing schema upgrades asynchronously, the main one is that most of the services expose a raw connection, in many cases for testing but in other cases those are exposed to other components or add-ons (unfortunately). Another problem is that most of the code doing those schema migrations has very weak error checking that depends on being synchronous.
So some of the components must be carefully rewritten in better async fashion to avoid that.
Regarding formHistory, looks like mxr is again behind and not updating for whatever reason, the code is much different after bug 566746, the situation should be quite better than reported.
MXR looks to be up to date to me for those files.
nsFormHistory.js is clearly still be using main thread/synchronous SQLite. But AIUI that's the "old" back-end that desktop isn't be using anymore (in favor of FormHistory.jsm). The only other commitTransaction call is in the db migration in FormHistory.jsm, and that one really shouldn't be showing up frequently.
nsIFormHistory2 still has consumers in metro, android and sync, though, and I also see a user in SafariProfileMigrator.js and in one test (browser_sanitizeDialog_treeView.js). Seems to me we need to at least get bugs on file to remove all of these uses and actually get rid of nsIFormHistory2.
I wonder whether the sync consumer (services/sync/modules/engines/forms.js) is responsible for the remaining slowsql issues...
(Also, do we need a bug on file to have FormHistory.jsm use SQLite.jsm instead of rolling its own?)
(In reply to :Gavin Sharp (use firstname.lastname@example.org for email) from comment #3)
> The only other commitTransaction call is in the
> db migration in FormHistory.jsm, and that one really shouldn't be showing up
...but we should still get rid of it. As far as I can tell there isn't much reason why FormHistory.jsm's DB migration couldn't be async.
Mark, can you chase this down further (comment 3/comment 4)?
Maybe this is moot if we can just remove all consumers and get rid of the old component entirely, but as it is these two components (nsFormHistory.js and FormHistory.jsm) seem to operate on the same database (formhistory.sqlite), and I wonder if that's going to cause any issues. E.g. there seems to be "migration" code in both of them...
As best as Gavin and I can tell, the slowsql dashboard is showing the number of "commit transaction" instances declining - and looking at some of the other related SQL statements, we believe there are still some older nightlies out there which haven't updated since bug 566746 landed. Once this stabilizes we should find the remaining cases are in other modules which continue to use nsIFormHistory2 - metro and sync.
So we'll treat this as a meta bug to remove nsIFormHistory2 and its implementation in nsFormHistory.js, which should kill those remaining cases. In the meantime, we can continue to monitor that dashboard to verify our assertions above as more nightlies update themself - at which point we can re-evaluate if this bug should keep its [snappy:p1] status.
Created attachment 757982 [details]
total number of sessions with fomrhistory slowsql
Created attachment 757983 [details]
number of slow sql queries(eg multiple queries per session)
Created attachment 757984 [details]
total number of sessions with formhistory slowsql
jydoop script used to generate the data: https://github.com/mozilla/jydoop/blob/master/scripts/slowsql.py
ran it with time make ARGS="scripts/slowsql.py outputfile 20130512 20130603" hadoop
The synchronous API should have no internal consumers in our codebase (apart dedicated tests) and warn deprecation to consumers at this point. If you can find anything else that is not a test, please file a dependency.
This bug will take care of finally removing the API and all remaining consumers, once we are comfortable with the removal, usually we wait at least a couple versions with the deprecation.
Data from bug 878677 comment 37 confirms that we've eliminated the majority of the form history main thread queries - good work everyone :)
(In reply to :Gavin Sharp (use email@example.com for email) from comment #12)
> Data from bug 878677 comment 37 confirms that we've eliminated the majority
> of the form history main thread queries - good work everyone :)
Note to myself since it's not obvious from bug metadata - this happened in the Firefox 24 cycle.
this is actionable.
Mark, are you actively working on this, or would you like to mentor the remaining work here?
Not working on it and very happy to mentor!
Cool, when you have the time, please post a comment or dependent bugs with leftover work to do.
So far I just did a quick search
Hi Mark! I would like to take this up. Can you please help me with this? :)
(In reply to Aman Dwivedi from comment #19)
> Hi Mark! I would like to take this up. Can you please help me with this? :)
Sorry for the delay and I'm really glad you want to help :)
If you look at the search link in comment 18, you will see there are a small number of references to nsIFormHistory2 left in the tree. Specifically:
* nsIFormHistory.idl can be removed - a reference to it will also need to be removed from toolkit/components/satchel/moz.build.
* The references to nsIFormHistory.h can be ignored - removing the IDL will take care of that.
* services/common/tests/unit/test_async_querySpinningly.js can also be removed - it's an old test that isn't relevant now that sync itself no longer uses this interface. There will be a reference to this file in xpcshell.ini that will also need to be removed.
* There is a reference in a comment in toolkit/components/satchel/FormHistory.jsm which can be left alone.
And as far as I can tell, that's it (although obviously we will be looking for a try run once the patch is up, but we can help you out with that.
Mak, does that sound right to you, or is there something I'm missing?
(In reply to Mark Hammond [:markh] from comment #20)
> * services/common/tests/unit/test_async_querySpinningly.js can also be
> removed - it's an old test that isn't relevant now that sync itself no
> longer uses this interface. There will be a reference to this file in
> xpcshell.ini that will also need to be removed.
To me looks like that method is still used in history and bookmarks engines:
And I don't see another test checking that querySpinningly works fine.
Maybe the test should be modified and simplified to use a memory database instead of form history?
> * There is a reference in a comment in
> toolkit/components/satchel/FormHistory.jsm which can be left alone.
I'd modify the comment to just state "// We don't support any schema version before FormHistory.jsm."
And nsFormHistory.js should also be removed, since it's the implementation of nsIFormHistory2. Thus also these references should be removed:
Created attachment 8826998 [details] [diff] [review]
(In reply to Marco Bonardo [::mak] from comment #21)
> (In reply to Mark Hammond [:markh] from comment #20)
> > * services/common/tests/unit/test_async_querySpinningly.js can also be
> > removed - it's an old test that isn't relevant now that sync itself no
> > longer uses this interface. There will be a reference to this file in
> > xpcshell.ini that will also need to be removed.
> To me looks like that method is still used in history and bookmarks engines:
> And I don't see another test checking that querySpinningly works fine.
Yeah, I think that's probably OK, but not ideal.
> Maybe the test should be modified and simplified to use a memory database
> instead of form history?
In this patch, I just did a few hacks allowing the test to still use the form history DB - it's a bit hacky, but IMO it is OK and exercising what it is trying to exercise. I expect querySpinningly to be dead later this year.
Aman, if you apply this patch to that problematic test and perform all the other steps described, I'd be happy to review it.