Watch the open files limit when running Riak

I was close to expressing my unbridled joy at how little hand-holding our Riak cluster needs, when we started to see strange increased latencies when hitting the cluster, on calls that should have been very fast. Also, the health of the Riak nodes seems fine in terms of CPU, memory and disk. As usual, our good old friend the error log file pointed us towards the solution. We saw entries like this in /var/log/riak/error.log:

2014-10-11 03:22:40.565 UTC [error] <0.12830.4607> CRASH REPORT Process <0.12830.4607> with 0 neighbours exited with reason: {error,accept_failed} in mochiweb_acceptor:init/3 line 34
2014-10-11 03:22:40.619 UTC [error] <0.168.0> {mochiweb_socket_server,310,{acceptor_error,{error,accept_failed}}}
2014-10-11 03:22:40.619 UTC [error] <0.12831.4607> application: mochiweb, "Accept failed error", "{error,emfile}"

A google search revealed that a possible cause of these errors is the dreaded open file descriptor limit, which is 1024 by default in Ubuntu.

To be perfectly honest, we hadn't done almost any tuning on our Riak cluster, because it had been running so smoothly. But recently we started to throw more traffic at it, hence issues with open file descriptors made sense. To fix it, we followed the advice in this Riak doc and created /etc/default/riak with the contents:

ulimit -n 65536

We also took the opportunity to apply the networking-related kernel tuning recommendations from this other Riak tuning doc and added these lines to /etc/sysctl.conf:

net.ipv4.tcp_max_syn_backlog = 40000
net.ipv4.tcp_timestamps = 0
net.ipv4.tcp_sack = 1
net.ipv4.tcp_window_scaling = 1
net.ipv4.tcp_fin_timeout = 15
net.ipv4.tcp_keepalive_intvl = 30
net.ipv4.tcp_tw_reuse = 1

Then we ran sysctl -p to update the above values in the kernel. Finally we restarted our Riak nodes one at a time.

I am happy to report that ever since, we've had absolutely no issues with our Riak cluster.  I should also say we are running Riak 1.3, and I understand that Riak 2.0 has better tests in place for avoiding this issue.

I do want to give kudos to Basho for an amazingly robust piece of technology, whose only fault is that it gets you into the habit of ignoring it because it just works!


Popular posts from this blog

Performance vs. load vs. stress testing

Running Gatling load tests in Docker containers via Jenkins

Dynamic DNS updates with nsupdate and BIND 9