When we moved to a write through cache architecture we also moved the cache write to offset based instead of APPEND.
We were inadvertently clearing our offset from our cache when we would clear which meant if the next operation was another write we would have the wrong offset and overwrite previous messages.
Signed-off-by: Derek Collison <derek@nats.io>
The original design had a shared filestore write buffer and individual message blocks had a read cache.
This presented some performance and stability issues when lots of reads and deletes were happening to a
message block that was also being written to actively.
This change eliminates the shared write buffer and uses the message block's cache as a write through as
well as read cache and handles partials correctly.
Signed-off-by: Derek Collison <derek@nats.io>
This will track the stream pending state for each consumer.
This code does account for filtered consumers.
Signed-off-by: Derek Collison <derek@nats.io>
Addresses stack overflow issue wally was seeing with configs
that mix and match streams and services between each other.
Signed-off-by: Derek Collison <derek@nats.io>
When creating shadow subscriptions for import streams, we were
not invoking code for gateway subscription accounting, which means
that when the account (for leafnodes) was switched to interest
only, those shadow subscriptions were not sent.
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
- Fix for updating delivery subject and adjusting next delivery sequences.
- When acking explicitly but out of order, need to make sure we set floor correctly.
- Only update ack floors on an ack if the message is present.
- Fix for needAck for explicitAck out of order consumers detecting if message has been acked.
- Fix for race not locking stream when checking interest during stop.
- Fix for filestore determing if a message block still has a message. Added check to first sequence as well as cache.
- Some additions to the original test.
Signed-off-by: Derek Collison <derek@nats.io>
The test shows the issue.
It seems that Consumer.needAck() for AckExplicit should consider
that an Ack is needed if sseq > o.asflr and there is no pending
ack at all. However making this change would break the test
TestJetStreamInterestRetentionStream.
Also, running the new test with `-count 100` and by adding an
artificial delay in stream.ackMsg() (just before calling
mset.store.RemoveMsg(seq)) causes sometimes delete requests
for the same sequence to be processed twice, which causes
the new test to fail (even with an attempted fix as discussed
above). I think that the attempt to remove the same sequence twice
is messing up the state.
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
The panic was caused by the closing of an already closed Go channel.
The Delete() relied on the consumer's mset being nil to consider
the consumer already closed. However, the consumer's mset is set
to nil after invoking sendDeleteAdvisoryLocked() which internally
invokes sendAdvisory() which releases/reacquires the consumer lock.
This left an open door for a race to occur and Delete() to be
invoked twice on the same consumer.
Moving setting the consumer's mset to nil too early would prevent
the sendAdvisory() to actually do its job. We could pass the mset
to sendAvisory(), but a simpler approach is to simply use a "closed"
boolean on the Consumer object that is set to true at the beginning
of the Delete() function.
Resolves#1621
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
It seems that when updating the delivery subject, we use as the
first sequence the ack floor, but if no message was ever ack'ed
then it causes the first message to be redelivered twice.
Resolves#1619
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
If the soliciting side detects the disconnect and attempts to
reconnect but the accepting side did not yet close the connection,
a "loop detected" error would be reported and the soliciting server
would not try to reconnect for 30 seconds.
Made a change so that the accepting server checks for existing
leafnode connection for the same server and same account, and if
it is found, close the "old" connection so it is replaced by
the "new" one.
Resolves#1606
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
This causes nil messages mid content stream and is confusing,
the test for this is amended to test the content received for
validity - where previously the test assumed any content is
good content but in fact it only received half the content.
Removing this behaviour until we can design this properly
Signed-off-by: R.I.Pienaar <rip@devco.net>
Users and NKey users will now have the option to specify a list
of allowed connection types.
This will allow for instance a certain user to be allowed to
connect as a standard NATS client, but not as Websocket, or
vice-versa.
This also fixes the websocket auth override. Indeed, with
the original behavior, the websocket users would have been bound
to $G, which would not work when there are accounts defined, since
when that is the case, no app can connect/bind to $G account.
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
Latency reports will include the header(s) responsible for the trace
Updated ADR to have it reflect implementation
Signed-off-by: Matthias Hanel <mh@synadia.com>
Made changes to processSub() to accept subscription properties,
including the icb callback so that it is set prior to add the
subscription to the account's sublist, which prevent races.
Fixed some other racy conditions, notably in addServiceImportSub()
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
