- Use stack buffers
- Ensure that buffer size is no greater than 90% of max_pending
- Added test with low max_pending
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
Use pending bytes as slow consumer trigger, so reintroduce max_pending.
Improve latency with inplace flush calls when appropriate. Utilize simple
time budget for readLoop routine.
Signed-off-by: Derek Collison <derek@nats.io>
When a route connection is created, the server will keep track
of the client structure in a special map until the route protocol
completes. This is meant so that if the server is shutdown before
the route is registered in routes map, the server can kick out
the connection's readLoop.
The route connection was correctly removed on success, but was
not for route connections that were not registered and dropped.
This was not causing any issue, but for correctness, doing the
removal now when server removes a route connection.
Signed-off-by: Ivan Kozlovic <ivan@synadia.com>
I noticed that when running the test suite, there would be a file
server/log1.txt left. This file is created by one of the config
reload test. Running this test individually was doing the proper
cleanup. I noticed that the Signal test that was checking
that files could be rotated was causing this side effect.
It turns out that none of the config reload tests were disabling
the signal handler (NoSigs=true), and since the go routine would
be left running, running the TestSignalToReOpenLogFile() test
would interact with an already finished test.
I put a thread dump in handleSignals() to track all tests that
were causing this function to start the go routine because NoSigs
was not set to true. I fixed all those tests. At this time, there
are only 2 tests that need to start the signal handler.
I have also fixed the code so that the signal handler routine select
on a server quitCh that is closed on shutdown so that this go routine
exit and is waiting on using the grWG wait group.
This PR is based out of #633. It imroves parsing QRSID so that the
TestRouteQueueSemantics test now passes (when dealing with malformed
QRSID).
A test similar to what is reported in #632 was also added. This
test however, uncovers a race condition that will be fixed in a
separate PR.
Resolves#632
This is the result of flapping tests in go-nats that were caused
by a defect (see PR https://github.com/nats-io/go-nats/pull/348).
However, during debugging, I realize that there were also things
that were not quite right in the server side. This change should
make it the notification of cluster topology changes to clients
more robust.
When a server accepts a route, it will keep track of that server
`connectURLs` array. However, if the server was creating a route
to that other server at the same time, it will promote the route
as a solicited one. The content of that array was not transfered,
which means that on a disconnect, it was possible that the cluster
topology change was not properly sent to clients.
Until now, a server would only notify clients of servers that join
the cluster. More than that, a server would send ot its clients only
information if new servers were added.
This PR changes this by sending to clients that support async INFO
the list of URLs for all servers in the cluster any time that there
is a change (joining or leaving the cluster).
As of now, clients will not be affected by the change (and will not
take benefit of this: removing servers from their server pool). This
will be addressed in each supported client once this is merged.
When A connects to B and B connects to A (either based on static
configuration - explicit routes, or because of auto-discovery -
implicit routes), it is possible that each server initially
registers the route from the opposite TCP connection. It will
then result in each server dropping the connection.
We were previously setting a retry flag in the first accepted route
based on the name of servers, which means that regardless of
duplicate detection, the server with the "smaller" server name would
try to reconnect when the route connection was closed. For instance,
suppose that server B connects to server A, when B disconnects, A
would try to reconnect once to B. This became problematic in the
case of configuration reload, because removing the route from B to
A would still result in a route created from A to B.
Also, when a route attempts a reconnect, a random delay is added
to avoid repeated failure cycles that may occur in case where
A connects to B and B to A.
This happens sometimes, and the latest occurence was today:
https://github.com/nats-io/java-nats/issues/96
When it happens, there is no error but subscribers would not receive
anything, etc...
This PR uses the fact that clients set the field Lang in the CONNECT
protocol that ROUTEs do not. I have checked that all Apcera supported
clients do set Lang in the CONNECT protocol.
If we plan to add Lang for routes, we need to find another field or
use a new one, in which case that would work only for new clients
(that would need to be updated).
With this change, when the server accepts a connection on the route
port and detects that this protocol field is present, it now closes
the client connection.
The nice thing is that newer clients, when incorrectly connecting
to the route port, get from the route's INFO the list of client URLs,
which means that on the initial connect error, they are able to
subsequently connect to the proper client port, so it is transparent
to the user (which may or may not be a good thing). However, it is not
guaranteed because if the client is not setting NoRandomize to true,
the client URL is added but the array shuffled, so it is possible that
the client library does not find the correct port in the connect loop.
The RunServer() function (and the various variants)
call Server.Start() in a go-routine, but do not return until
it has verified that the server is ready to accept connections.
To do so, it use GetListenEndpoint() to get a suitable connect
address (replacing "0.0.0.0" or "::" with localhost - important
on Windows). It then creates a raw TCP connection to ensure the
server is started, repeating the process in case of failure up
to 10 seconds.
This PR replaces this with a function that checks that client
listener, and route listener if configured, are set. This removes
the need to get a connect address and create test tcp connections.
The reason for this change is that NATS Streaming when starting
the NATS Server (unless configured to connect to a remote one)
calls RunServerWithAuth(), which when getting "localhost" from
GetListenEndpoint(), would fail trying to resolve it. This happened
for the NATS Streaming Docker image built with Go 1.7+.
Trying to use IPv6 address for the cluster host would fail.
Also, there were some unclosed channels in case of accept loop
setup failures.
Resolves#323
