While I could use memcached on Debian to use the default 11211 port, but I've had great difficulty setting up Unix socket.
From what I've read, I know that I need to create a memcache.socket
and add:
-s /path/to/memcache.socket
-a 0766
To /etc/memcached.conf
and comment out the default connection port and IP, i.e.
-p 11211
-l 127.0.0.1
However, when I restart memcached I get internal server errors on Drupal site.
I'm trying to implement unix sockets to avoid TCP/IP overhead and boost overall memcached performance, however not sure how much performance gain one can expect of this tweak.
I appreciate your hints or possibly configs to to resolve this.
You may find that just setting the socket path doesn't work. Memcached drops privileges before it creates its socket, though after it's written its PID. It's common to put the socket in
/var/run
(e.g. as mysql does), but only root can write there, so create/var/run/memcached
and chown it to nobody, then set/var/run/memcached/memcached.sock
as the socket path. Writing it to/tmp
would work too, but by its very nature that could get trashed; sticking it in/var/run
is a bit more permanent.You can check it's working by using netcat to connect to it:
nc -U /var/run/memcached/memcached.sock
Just type
stats
at the blank prompt; if it's working you'll get a load of output.I believe the socket will need to be world executable, not writable. If your seeing a PHP notice message containing errno=32 Broken pipe, adjust Memcached access mask to 0755
Also verify that Drupal's Memcached daemon has TCP port of
0
in the configuration. An unset port will default to11211
, and confuse the socket connection.added to /etc/default/memcached:
added to /etc/memcached.conf:
I benchmarked the performance boost of using memcached unix sockets, it is 33% faster to use unix sockets in my tests using the php cli.
I found that it is important to add the memcache user to the www-data group and give the unix socket 775 permissions, this way both the php handler and memcache user can execute it.
You can find the benchmarks and write-up here - although it is for WordPress it should work for Drupal as it is very likely a permissions issue.
Note if you are using an alpine container in Kubernetes the pod spec with health probes would look something like this:
Took a while to figure out the right settings, and how to connect to a unix local domain socket via busybox.