A message block is checking the filestore's cfg.Subjects to see
if it can "intern" the subject or not. The problem is that this
is done under the message block's lock, but not the filestore.
However, during a stream configuration update, the filestore's
cfg field is switched to a new one, causing the datarace.
By making sure we do the switch under all message blocks lock,
we remove the data race (that could be reproduce by running th
test TestJetStreamClusterMoveCancel with -count=10).
We investigating the use of a string interning library but it
showed a little performance degradation that this approach does
not suffer from.
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
Tests that run on Travis have been split into jobs that run in their own VM in parallel. This reduces the overall running time but also is allowing recycling of a job when we get a flapper as opposed to have to recycle the whole test suite.
JetStream Tests
For JetStream tests, we need to observe a naming convention so that no tests are omitted when running on Travis.
The script runTestsOnTravis.sh will run a given job based on the definition found in ".travis.yml".
As for the naming convention:
All JetStream tests name should start with TestJetStream
Cluster tests should go into jetstream_cluster_test.go and start with TestJetStreamCluster
Super-cluster tests should go into jetstream_super_cluster_test.go and start with TestJetStreamSuperCluster
Not following this convention means that some tests may not be executed on Travis.
880e854637