[ZODB-Dev] transaction timeouts raise no exceptions and cause strange errors on clients

[Jim Fulton]

On Thu, Nov 4, 2010 at 10:31 AM, Chris Withers <chris at simplistix.co.uk> wrote:
> On 04/11/2010 13:43, Jim Fulton wrote:
>>> I'm curious:
>>>
>>> - Why is the timeout not logged at ERROR or even CRITICAL?
>>
>> Because it was considered to be neither.  It's also not a server error.
>
> It is an erroneous condition; a transaction has taken long enough that a
> timeout has occurred. Thinking about it, it feels more like a
> TimeoutException should be raised, but then would the storage server
> clean things up properly?
>
>>> - Is the "inconsistent state" warning genuine here?
>>
>> Yes.  You don't really know whether the transaction committed or not.
>> It's much worse if multiple storages are involved.
>
> ...which of course, they are ;-)
>
>>> - If that warning is genuine, is there any way the timeout could not
>>>    leave the client in an inconsistent state?
>>
>> Absolutely.
>
> My wording was poor:

So was my reading. (I missed the "not".)

> If that warning is genuine, is there any way the timeout could be made
> to not leave the client in an inconsistent state?

No.

>
>> This is why I tend to view transaction timeouts as a last resort and
>> set the limit
>> much higher, 5 minutes.
>
> Fair enough. Seems odd to have a feature which you seem to be suggestion
> should not be used...

Actually, I think it should be used, at least for applications with
high availability requirements. For applications like our's (ZC's)
it amounts to a "don't be stupid" setting.

For some applications, it's better to risk inconsistency than to have
your application grind to a halt.  I consider it an emergency device
though, which is why I'd set it pretty high.

To be clear, you never want to time out transactions that aren't
already hosed.

Jim

-- 
Jim Fulton