Ben Turner (not reading bugmail, use the needinfo flag!) Assignee
	Description • 10 years ago

The proposal is to change our IndexedDB write transactions to more closely align with Chrome/IE which favor performance over durability. We will have an opt-in mechanism to have per-transaction durability for consumers that care.

	Jan Varga [:janv]
	Comment 1 • 10 years ago

+1

	Andrew Sutherland [:asuth] (he/him)
	Comment 2 • 10 years ago

Can you clarify the risk profile here?  Specifically, a crash shortly-after a non-durable transaction commits may result in:
A) Database corruption.  The whole database may need to be blown away because the pages of the database may not be in a coherent state. (non-WAL PRAGMA SYNCHRONOUS=0/OFF, low probability on non-WAL PRAGMA SYNCHRONOUS=1/NORMAL)
B) Loss of the most recent transaction but the database should probably be coherent. (non-WAL PRAGMA SYNCHRONOUS=1/NORMAL)
C) Loss of multiple recent transactions but the database will still be coherent at that earlier point.  (WAL with SYNCHRONOUS=1/NORMAL and where a checkpoint is not triggered)

Since IndexedDB currently uses PRAGMA synchronous=2/FULL, I'm assuming this is a proposal to drop to NORMAL or OFF?  Or is the idea to switch to using WAL with PRAGMA synchronous=1/NORMAL and the opt-in mechanism would just make sure to trigger a checkpoint as part of the transaction?

The WAL option sorta sounds preferable since the non-WAL options seem more likely to result in database corruption, requiring apps to make every transaction be durable.  Or to shard things so that there is one IndexedDB database that is where the really important stuff is (config, drafts for email), and another happy-go-lucky DB for stuff that isn't authoritative (synced email).

	Ben Turner (not reading bugmail, use the needinfo flag!) Assignee
	Comment 3 • 10 years ago

IDB currently runs with non-WAL and SYNCHRONOUS=NORMAL (unless you set an obscure pref that might flip it to FULL).

The change will be to make normal write transactions run with WAL and SYNCHRONOUS=NORMAL. If WAL mode isn't available for some reason then we'll still use SYNCHRONOUS=NORMAL.

The specially-opted-into durable write transactions will essentially run with WAL and SYNCHRONOUS=FULL (it won't actually because we're changing the way we use connections, but that'll be the end result).

	Andrew Sutherland [:asuth] (he/him)
	Comment 4 • 10 years ago

Awesome.  Thanks for the clarification!

	Kyle Huey (Exited; not receiving bugmail, old account, do not use)
	Comment 5 • 10 years ago

We're absolutely not going to introduce opportunities for database corruption.

	Ben Turner (not reading bugmail, use the needinfo flag!) Assignee
	Comment 6 • 10 years ago

(In reply to ben turner [:bent] (use the needinfo? flag!) from comment #3)
> (it won't actually because we're changing the way we use connections,
> but that'll be the end result).

Through manual checkpointing.

	Ben Turner (not reading bugmail, use the needinfo flag!) Assignee
	Comment 7 • 9 years ago

This gives us an optional mode to say "no, really, I want it on disk".

	Kyle Huey (Exited; not receiving bugmail, old account, do not use)
	Comment 8 • 9 years ago

Comment on attachment 8562320 [details] [diff] [review]
Patch, v1

Review of attachment 8562320 [details] [diff] [review]:
-----------------------------------------------------------------

We should test that if experimental features are not enabled then readwriteflush throws.

r=me w/that

	Ben Turner (not reading bugmail, use the needinfo flag!) Assignee
	Comment 9 • 9 years ago

https://hg.mozilla.org/integration/mozilla-inbound/rev/0b4de21b759f

	Phil Ringnalda (:philor)
	Comment 10 • 9 years ago

Backed out in https://hg.mozilla.org/integration/mozilla-inbound/rev/beb016173f6f - see bug 866846 comment 21

	Ben Turner (not reading bugmail, use the needinfo flag!) Assignee
	Comment 11 • 9 years ago

https://hg.mozilla.org/integration/mozilla-inbound/rev/3be650e84eea

	Wes Kocher (:KWierso) (Not reading bugmail; email directly if needed)
	Comment 12 • 9 years ago

Backed out for frequent bc1 oranges
https://treeherder.mozilla.org/logviewer.html#?job_id=8256324&repo=mozilla-inbound

https://hg.mozilla.org/integration/mozilla-inbound/rev/0b12fc331918

	Ben Turner (not reading bugmail, use the needinfo flag!) Assignee
	Comment 13 • 9 years ago

https://hg.mozilla.org/integration/mozilla-inbound/rev/b0c5f21c7811

	Ryan VanderMeulen [:RyanVM]
	Comment 14 • 9 years ago

https://hg.mozilla.org/mozilla-central/rev/b0c5f21c7811

	Lawrence Mandel [:lmandel] (use needinfo)
	Comment 15 • 9 years ago

Release Note Request (optional, but appreciated)
[Why is this notable]: This will have an impact on devs.
[Suggested wording]: IndexedDB transactions now non-durable by default
[Links (documentation, blog post, etc)]:

	Kohei Yoshino [:kohei]
	Comment 16 • 9 years ago

Is there any examples that will be affected by this change? I see the new readwriteflush mode in the patch but readwrite is still the default mode, right?

	Ben Turner (not reading bugmail, use the needinfo flag!) Assignee
	Comment 17 • 9 years ago

(In reply to Kohei Yoshino [:kohei] from comment #16)
> Is there any examples that will be affected by this change? I see the new
> readwriteflush mode in the patch but readwrite is still the default mode,
> right?

Well, the default mode for transactions is actually 'readonly'. But the new 'readwriteflush' mode is not available unless you have enabled the indexedDB experimental feature preference. Without that preference set you are only able to use 'readonly' and 'readwrite'.

	Kohei Yoshino [:kohei]
	Comment 18 • 9 years ago

Oh I meant 'readonly', of course I know that because my app uses IndexedDB.

> optional IDBTransactionMode mode = "readonly"

So I guess there would be no backward compatibility issues here.

	Ben Turner (not reading bugmail, use the needinfo flag!) Assignee
	Comment 19 • 9 years ago

(In reply to Kohei Yoshino [:kohei] from comment #18)
> So I guess there would be no backward compatibility issues here.

Right.

	Kohei Yoshino [:kohei]
	Comment 20 • 9 years ago

Thanks! Removing the relnote-firefox flag given the dom.indexedDB.experimental pref is defaulted to false at this moment.

	Ben Turner (not reading bugmail, use the needinfo flag!) Assignee
	Comment 21 • 9 years ago

(In reply to Kohei Yoshino [:kohei] from comment #20)
> Thanks! Removing the relnote-firefox flag given the
> dom.indexedDB.experimental pref is defaulted to false at this moment.

Well, this change is still worthy of a relnote in my opinion (and see comment 15). The change won't affect the code that people write but it will affect the timing of the 'complete' event and the behavior in the event of a system crash.

	Kohei Yoshino [:kohei]
	Comment 22 • 9 years ago

Ah, then put the relnote flag back... I don't add the site-compat flag though.

	Chris Mills [:cmills]
	Comment 23 • 9 years ago

I've written up a definition of durability and what the current state in browsers is:

https://developer.mozilla.org/en-US/Firefox/Releases/40#IndexedDB

I've also mentioned this on relevant pages:

https://developer.mozilla.org/en-US/docs/Web/API/IDBDatabase/transaction
https://developer.mozilla.org/en-US/docs/Web/API/IDBTransaction/oncomplete

And added a link to the definition in the release notes:

https://developer.mozilla.org/en-US/Firefox/Releases/40#IndexedDB

bent helped me write this, but another tech review would be cool, just as a sanity check.

	Ben Turner (not reading bugmail, use the needinfo flag!) Assignee
	Comment 24 • 9 years ago

Thanks Chris! One thing to note though is that the "readwriteflush" mode is only available if the "dom.indexedDB.experimental" pref is set to true.

	Chris Mills [:cmills]
	Comment 25 • 9 years ago

Added this info; thx Ben.

And the correct link for the 'durable' definition is:

https://developer.mozilla.org/en-US/docs/Web/API/IndexedDB_API/Basic_Concepts_Behind_IndexedDB#durable