First, the test should be done only for the initial INFO and only
for solicited connections. Based on the content of INFO coming
from different "listen ports", use the CID and LeafNodeURLs for
the indication that we are connected to the proper port.
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
When the server logs information related to a connection, it uses
the connection IP and remote port as a prefix. When it was an IPv6
address, the square brackets would be missing.
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
Also make the wait bound to 3secs after which writeLoop will attempt
to flush. Will log if it timed out on the wait and entering with
fsp > 0. Limit the report to once every 10 minutes
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
This related to PR #1233.
The computation of the time to stall a fast producer was bogus. Fixed
that and added a unit test for the function computing this stalled
duration.
Also, in PR #1233, I had removed Gosched() when a call to flushOutbound()
realizes that the flag is already set. It was forgetting that readLoop
in some cases will call flushOutbound() in place. So there is still
value in unlock/gosched/lock again in that function.
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
When an account is switched to interest-only mode due to no interest,
it was not possible to switch that account more than once. But the
function switchAccountToInterestMode() that triggers a switch could
possibly doing it more than once. This should not cause problems
but increased the number of traces in a big super cluster.
Also fixed some flappers and a data race.
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
- All writes will now be done by the writeLoop, unless when the
writeLoop has not been started yet (likely in connection init).
- Slow consumers for non CLIENT connections will be reported but
not failed. The idea is that routes, gateway, etc.. connections
should stay connected as much as possible. However if a flush
operation times out and no data at all has been written, the
connection will be closed (regardless of type).
- Slow consumers due to max pending is only for CLIENT connections.
This allows sending of SUBs through routes, etc.. to not have
to be chunked.
- The backpressure to CLIENT connections is increased (up to 1sec)
based on the sub's connection pending bytes level.
- Connection is flushed on close from the writeLoop as to not block
the "fast path".
Some tests have been fixed and adapted since now closeConnection()
is not flushing/closing/removing connection in place.
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
Updated all tests that use "async" clients.
- start the writeLoop (this is in preparation for changes in the
server that will not do send-in-place for some protocols, such
as PING, etc..)
- Added missing defers in several tests
- fixed an issue in client.go where test was wrong possibly causing
a panic.
- Had to skip a test for now since it would fail without server code
change.
The next step will be ensure that all protocols are sent through
the writeLoop and that the data is properly flushed on close (important
for -ERR for instance).
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
- Call flushOutbound() for SYSTEM connections
- Flush in place in internalSendLoop when sending the shutdown event
- Fix some tests:
- missing defer client connection Close()
- ensure subs are registered and messages received before shutdown
of leafnode server to check disconnected event's stats.
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
This could happen if the remote server is running but not dequeueing
from the socket. TLS connection Close() may send/read and so we
need to protect with a deadline.
For non client/leaf connection, do not call flushOutbound().
Set the write deadline regardless of handshakeComplete flag, and
set it to a low value.
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
Running test suite on a Windows VM, I notice several failures.
Updated the compute of the RTT to be at least 1ns. I think that
this is just an issue with the VM I am running, but that change
will have no impact for normal situations (since setting the rtt
to the very minimum duration (1ns) instead of 0) and will prevent
some tests from failing.
Because of those same timer granularity issues, I had to add some
delays between some actions in order for time.Sub()/Since() to
actually report something more than 0.
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
We had too much special processing, so reduced to a single wildcard
which will propagate across routes and gateways and is consistent
with gateway handling of globally routed subjects and timeouts.
Signed-off-by: Derek Collison <derek@nats.io>
Use centralized sync map to gather *client that have GW replies.
Tested with concurrent receiving clients and perf is as good as
with timer per client but reduces need of that timer per client
object.
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
- Add atomic in client to skip check in processInboundClientMsg()
if value is 0. Avoids getting the lock in fast path if not needed.
- Have a timer per client instead of the global server list that
was expiring: noticed a lot of contention there when running
some perf/profiling tests. The timer is also not reset for
every timestamp that is not yet expired since this too affects
performance. Instead fires are regular interval and cleared
when map is empty after a cycle.
- Move processing of gw map rely on its own function (in inbound msg).
I have verified that this is inlined same way as when code was
directly in processInboundClientMsg.
- Use string(subj[]) for prefix detection: I have verified that
it is actually faster.
- Builds the RMSG with appends to local buffer in handleGatewayReply()
instead of using fmt.Sprintf().
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
- New prefix that includes origin server for the request
- Mapping done if request is service import or requestor has
recent subscription
- Subscription considered recent if less than 250ms
- Destination server strip GW prefix before giving to client
and restore when getting a reply on that subject
- Mapping removed aftert 250ms
- Server rejects client publish on "$GNR." (the new prefix)
- Cluster and server hash are now 8 chars long and from base 62
alphabets
- Mapped replies need to be sent to leafnode servers due to race
(cluster B sends RS+ on GW inbound then RMSG on outbound, the
RS+ may be processed later and cluster A may have given message
to LN before RS+ on reply subject. So LN needs to accept the
mapped reply but will strip to give to client and reassemble
before sending it back)
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
Ivan had the idea of using the CONNECT to establish a first estimate of RTT
without additional PING/PONGs.
Signed-off-by: Derek Collison <derek@nats.io>
If a client RTT for a requestor is longer than a service RTT, the requestor latency was often zero.
We now wait for the RTT (if zero) before sending out the metric.
Signed-off-by: Derek Collison <derek@nats.io>
As soon as server has processed a client CONNECT, it was possible
that if Connz() or other was requested, the server will send a
PING to compute the RTT. This would cause clients that expect
the first PONG as part of synchronous CONNECT logic to fail.
Make sure that we delay the first RTT ping to after sending the
first PONG, or if client does not send PING as part of the CONNECT,
after 2 seconds have elapsed since the tcp connection was accepted.
Resolves#1174
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
Currently, the config file doesn't recognize the latency config block in
account exports. This change exposes those settings in the config file.
Signed-off-by: Jaime Piña <jaime@synadia.com>
Signed-off-by: Waldemar Quevedo <wally@synadia.com>
This is the first pass at introducing exported services to the system account for generally debugging of blackbox systems.
The first service reports number of subscribers for a given subject. The payload of the request is the subject, and optional queue group, and can contain wildcards.
Signed-off-by: Derek Collison <derek@nats.io>
Will now breakout the internal NATS latency to show requestor client RTT, responder client RTT and any internal latency caused by hopping between servers, etc.
Signed-off-by: Derek Collison <derek@nats.io>
Changed some of client.outbound fields to int64.
Moved fields around to minimize size of struct (checked with
unsafe.Sizeof())
Checked benchmark results before/after
Added test
Resolves#1118
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
With Go 1.12 (strangely was not able to reproduce with Go 1.11)
the test TestRouteNoCrashOnAddingSubToRoute() would frequently
locks up and consume all avail CPUs on the machine. Running
this test with GOMAXPROCS=2 you would see server.test CPU usage
pegged at 200% (assuming you have at least 2 CPUs).
The reason was that the writeLoop was spinning because another
routine was already in flushOutbound() and stack trace would
show that it was stuck in system calls. It seems that even though
the writeLoop does release the lock but grab it right away was
not allowing the syscall to complete.
So decided to put back the unlock/gosched/lock back in flushOutbound()
when flag is already set, but then protect the closeConnection()
with its own flag (similar to clearConnection) to not re-introduce
issue fixed in #1092.
Had to fix the benchmark test RoutedInterestGraph because after a
route is accepted, the initial PING will be sent after 1sec which
was breaking this test.
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
I noticed that TestNoRaceRoutedQueueAutoUnsubscribe started to
fail a lot on Travis. Running locally I could see a 45 to 50%
failures. After investigation I realized that the issue was that
we have wrongly re-used `subscription.nm` and set to -1 on unsubscribe
however, I believe that it was possible that when subscription was
closed, the server may have already picked that consumer for a delivery
which then causes nm==-1 to be bumped to 0, which was wrong.
Commenting out the subscription.close() that sets nm to -1, I could
not get the test to fail on macOS but would still get 7% failure on
Linux VM. Adding the check to see if sub is closed in deliverMsg()
completely erase the failures, even on Linux VM.
We could still use `nm` set to -1 but check on deliverMsg(), the
same way I use the closed int32 now.
Fixed some flappers.
Updated .travis.yml to failfast if one of the command in the
`script` fails. User `set -e` and `set +e` as recommended in
https://github.com/travis-ci/travis-ci/issues/1066
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
This was introduced in PR#930. The first commit had the route's
check if the flushOutbound() returned false, and if so would
locally unlock/lock the connection's lock. Unfortunately, this
was replaced in the second commit (a6aeed3a6b)
to the flushOutbound() function itself.
This causes the function closeConnection() to possibly unlock
the connection while calling flushOutbound(), which if the
connection is closed due to both a tls timeout for instance
and explicitly, it would result in the connection being scheduled
for a reconnect (if explicit gateway connection, possibly route).
Added defensive code in Gateway to register a unique outbound gateway.
Fixed a test that was now failing with newer Go version in which
they fixed url.Parse()
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
Take into account tracking of response maps that are created and do proper cleanup.
Also fixes#1089 which was discovered while working on this.
Signed-off-by: Derek Collison <derek@nats.io>
When a route receives a message, it uses a thread local cache to
find the account and subscriptions match for a given subject.
When not found, an entry is added to this cache. The problem is
that this cache will reference subscriptions that in turn
reference connections.
When the subscriptions/connections are closed, this thread local
cannot be purged from those closed subscriptions (since it is
thread local - no lockin is used).
The real issue is that connection's buffer was not set to nil on
close, which then could cause more than expected memory to be
still referenced. Setting the buffer to nil will help reduce the
memory being used.
When an entry is added to the cache, the cache may reach a size
that will cause the server to prune some entries. From time to
time, the cache will be scanned to look for entries that contain
only closed subscriptions and remove those.
Resolves#1082
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
