[ZODB-Dev] Potential BTrees splitting bug

[Tim Peters]

[Tim Peters]
>> I wonder whether running on a 2-CPU
>> box introduces a wrinkle we haven't seen before (I've never run the
>> tests on a 2-headed box -- don't have one)?

[Paul Winkler]
> Interesting. You mentioned earlier that a recent round of ZEO cache
> bugs was very difficult to provoke on Redhat linux.  But at my
> company we were constantly seeing stale data via ZEO, all on red hat
> boxes. The ZEO clients were all running on smp (intel) hardware.
>
> Does pythonlabs have no SMP box available for testing?

We do, but they're in test clusters that aren't easy to get at for
day-to-day use.  The desktops and laptops I can get to without hassles are
all single-CPU.  Curiously, all along "the best" OS for provoking timing
bugs has been Win98SE, because its thread and process scheduling is
inconsistent and seemingly incomprehensible.

> then again, my cache problems were apparently cured by moving to
> ZEO from ZODB-3.1.2 so maybe you're talking about a more
> recent bug.

If you think the last critical ZEO bug was fixed in 3.1.2, you better sit
down before reading the NEWS file for 3.1.3 <wink>:

    http://www.zope.org/Products/ZODB3.1/NEWS

Observation:  As the 3.1 line goes on, the bugs remaining get ever rarer in
practice, and ever harder to track down.  It's quite possible you've never
experienced the ZEO bugs fixed in 3.1.3; they happen to have been first
diagnosed on a sluggish multi-CPU 64-bit Solaris box under extreme load, in
a system configuration more complicated than any of the standard ZEO tests
approximate.