I would say the short answer is that it was a short ad-hoc test
task.
I don't think we ever had anything very formal. In the beginning
there were some store tests that did "crash recovery", by simply
doing a set of sql tests in order in ij and not exiting cleanly.
As each subsequent test was run on same db we would at least run through the
code path at boot time that looks at the log and recovers it.
More recently some utilities were added to make it easy to convert/write
new tests that exercise this path. For instance OCRecoveryTest uses
these utilities to do some work in a forked jvm, exit that JVM, and
then verify results on reboot.
Exiting the JVM tests most of what I would consider crash recovery, lets
say 99% or more. But the unflushed data is still in the computer cache.
So especially now that it is relatively easy to write
these kinds of tests in junit, time spent here would be best. Extreme
crash recovery would actually detach the disk before data could be
sync'd and turn off the machine. I think we may have done something
like this in early days and found some obvious bugs, but
would not expect much, as it is hard to look for correct results other
than a sucessful boot.
On 8/20/2014 11:14 AM, Rick Hillegas wrote:
I would like to understand what additional testing of the release
candidate should be performed in order to satisfy the checklist item
called "Recovery after crashing engine". Various recovery tests are run
as part of our platform release tests, but it I don't see any tests for
crash recovery.
We do have a crash recovery test for encrypted databases
(CryptoCrashRecovery), but that only runs on debug jars. So it has only
been exercised by nightly/continuousIntegration tests on the 10.11
branch, not on the release candidate itself.
I have some questions for the Store experts:
1) Does "Recovery after crashing engine" mean running
CryptoCrashRecovery on the debug distribution of 10.11.1.1?
2) Does it entail some other, formal tests?
3) Or is it meant to cover some other ad-hoc tests which are run by hand
and are not checked in under subversion?
Thanks,
-Rick
.
