It was fun playing with you and now Vic is following you in leaving the game.

Hope to see you come back, I returned two weeks ago, and it’s actually very nice to have no army to think about and be able to build the empire again up from the ruins

kind regards
-Har

We did however just have to write a backup for memcached…

We have a site which for client requirements reasons absolutely has to run off Oracle, but the site itself makes a huge amount of queries on the homepage. If a change is made and the homepage cache has to be emptied the load put on Oracle by the volume of traffic hitting the page takes down the Oracle server. To get around this, we run a queue which replaces keys in Memcached in the background, rather than deleting the cache and letting a user repopulate it with a hit.

Our problem though is if one of the memcached servers goes down for any reason such as power failure or any general failure. Because the users no longer rebuild the cache, there is no way to rebuild the lost content. Instead what we are doing is piping the memcached log through a deamon we have written which scans for PUTs and when a PUT is done it does a GET on that key and saves it out to an SQLlite database. This effectively gives us a background running mirror of each memcached instance. Then on boot, the contents of the SQLlite table is loaded back into memcached very quickly.

An ugly and unfortunate but effective solution to a painful problem.

The developer says that memcached might have some issues and there is nothing he can do about it.

Any suggestions on what to do?

Thanks

I do not delete cache entries and don’t need to understand what particular key stands for. Cache entries have short expire of 60 seconds and that’s enough in terms of deletion. All I am concerned of is performance = how much I could gain riping off md5 and what to chose instead. Thanks

Its also much easier to perform a delete against a key like that in another part of your code. Unless you have a method for every key you create, which is kinda overboard.

Not to mention, it would also be much harder to understand what user_c4ca4238a0b923820dcc509a6f75849b is vs user_c81e728d9d4c2f636f067f89cc14862c when looking at memcached -vv or other debug tools.

It takes a little more work, but its all about key organization.

My keys need to include method name + arguments passed, so I am using something like:
md5( serialize( array( method name, array( arguments passed ) ) ) ) as memcache key.

This keeps my keys unique and short (as serialized arguments might be longish sometimes). What would you suggest instead of md5 in that case?