Mario Juric created UIMA-6273:
---------------------------------
Summary: isBeforeV3 variable not reset in BinaryCasSerDes
Key: UIMA-6273
URL: https://issues.apache.org/jira/browse/UIMA-6273
Project: UIMA
Issue Type: Bug
Components: uimaj
Affects Versions: 3.1.1SDK, 3.1.0SDK, 3.0.2SDK, 3.0.0SDK, 3.0.1SDK
Reporter: Mario Juric
This issue that took some time to track down, but I think we finally managed to
identify the problem. It seems that the variable isBeforeV3 was introduced with
UIMA3, but it is never cleared when the CAS is reset, i.e. the flag is not set
to false again in the BinaryCasSerDes clear method. This means if the CAS
instances are recycled, as it is the case with the CPE, that the value will
never be false again after it has once been set to true.
We ran into the issue with some integration tests that verified loading of our
older CAS binaries, and they worked well until we introduced an UIMA 3 version
of the CAS binary into the tests. The error was strange because it seemed a bit
random at first without being random, until we discovered that the error didn’t
show up if the CAS was loaded within the first 3 tests. Turns out that the CPE
recycles 3 CAS’es when a single thread is active, so if the test ran as one of
the first 3 then it would start with a fresh CAS where isBeforeV3 is always
false. Once it ran as number 4 or later then isBeforeV3 would be true, because
all other tests were CAS binaries created under UIMA 2. The strange thing was
that it never caused any exception, it just loaded the CAS with most
annotations missing, so this added to the initial confusion.
Resetting isBeforeV3 in the clear method solves the issue for our part, but it
also requires that the variable is initialised later in the reinit method of
BinaryCasSerDes, since it could be reset to false when the delta flag is false,
and this can also cause failures during loading of earlier CAS versions than
v3. The isBeforeV3 flag is not accessed before deserializeAfterVersion is
invoked, and this only occurs inside this reinit method, so this additional
change should not cause any unwanted side effects.
I am not sure if there are other variables that need to be reset when the CAS
is reset, but I think it may be necessary to go through them to see if
something else has been missed. This issue affects all versions since at least
UIMA 3 was released.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
