Support multiple memcache daemons
jjneely opened this issue · comments
The original implementation supported multiple memcache servers. For my usage, running multiple memcached daemons per machine scales better than one and I was using that feature.
Late to the commit party, sorry....
This goes against the best practices in Prometheus as it adds several additional failure domains (a failure to retrieve stats from one node will result in no metrics for any server, dependency on the network, etc.).
It actually scales worse than having one exporter per node. The prometheus server is engineered to scrape from many targets, the exporters are usually not. With kubernetes(pods!)/marathon etc. or any kind of configuration management it's very easy to always scale two processes together.
On the bright side, you can now get memcached_process_* metrics for free (which really only works if exporter+memcached are on the same machine).
All the memcached daemons are on the same machine. We have a tier of machines that run three memcached daemons each.
I do understand the point, I just got everything working with the old implementation -- unaware that this was coming. I get Issue #2 to deal with too.
Well, you can of course continue to use the old version*. I expected some push back around that removal, which was part of the reason I kind of forked it into a memcached_exporter.
Scraping 3 instances on the same node is a bit of an edge case, at least there is no network involved. Though, the point remains that a failure to connect to any of the 3 memcached will result in a metric failure for all of them.
* Just be aware that the metrics are both incomplete (no metrics for set commands for example) and don't follow the prometheus naming.
The original implementation supported multiple memcache servers. For my usage, running multiple memcached daemons per machine scales better than one and I was using that feature.
Late to the commit party, sorry....
Since memcached does not support multiple databases, this would be really useful to be still implemented.