On Mon, Aug 16, 2010 at 3:06 AM, Jeremy Orlow jor...@chromium.org wrote:
On Fri, Aug 13, 2010 at 7:30 PM, Pablo Castro pablo.cas...@microsoft.com
wrote:
In the context of transactions, readers using READ_ONLY and writers using
READ_WRITE may block each other when starting transactions, at least for
cases where the underlying implementation uses locking for isolation. Since
we allow multiple readers and they can start while other readers were
already running, it's possible that readers end up starving writers in a
concurrent setting. It seems it would be a good idea to add some minimum
guarantees to the spec that ensures some amount of fairness to concurrent
activities against a given database.
We could either include a loose recommendation or try to mandate a strict
behavior. It seems the loose recommendation is more practical, the questions
are a) is there a risk of incompatible behavior because of
under-specification, and b) will we risk that some implementations will just
ignore this aspect if it's specified too informally.
The loose recommendation could just be a sentence in the transactions
section:
UAs need to ensure a reasonable level of fairness across readers and
writers to prevent starvation.
If we wanted to be more specific, we could go with something like this
(we'd probably spell it out as rules if we decide to put this strict version
in the spec):
All readers can run concurrently, but once a writer tries to start a
transaction we stop allowing new readers to start and queue up the writer
and any subsequent reader/writer. Once the existing readers are drained the
writer runs, and after that whatever is queued up next runs, which can be
another writer or all the remaining readers (depending upon what came first,
another writer or another reader; readers are released all simultaneously
since they run concurrently).
Given that not all implementations will have to deal with this and that
different implementations may want to have different strategies, it seems
that just having the recommendation around starvation is the best option.
I agree that we should have some (non-normative?) recommendation to avoid
starvation in the spec and that anything more formal would be too
constraining.
Sounds good to me. Here is the strategy that we're currently using in Firefox:
1. Transactions are always executed in the order they are started.
Unless they can execute in parallel.
2. Transactions are considered to be started when the first request is
placed against them.
So in the following code:
trans1 = db.transaction([foo]);
trans1.objectStore(foo).get(...).onsuccess = ...;
trans2 = db.transaction([foo], READ_WRITE);
trans2.objectStore(foo).put(...).onsuccess = ...;
trans3 = db.transaction([foo]);
trans3.objectStore(foo).get(...).onsuccess = ...;
trans3 won't start until trans2 is committed.
/ Jonas