Scaling Memcache at Facebook (2021)

Redis is simply more useful if you're not pushing max throughput using it purely as a cache. It has really useful support for a variety of data structures and operations, and I'd guess that Memcached's (previous) 1mb limit per item drove a good number of people to look for other options over time.

Eons ago now (circa 2009/2010 I think), when working on Pornhub we switched from memcached to Redis for a few reasons:

- memcache had little to no observability, it made debugging cache issues a nightmare. Redis on the other hand is a dream to work with on that front, from the cli that is well thought out to the fact you can copy the db file to your machine and go at it, it's just a much better developer experience.

- Data structures, Redis' lists, hashes and sets meant we could represent a lot of a stuff as is in cache with no need for costly serializing. It also meant the data layer code was simpler than the memcached equivalent.

- Built in pub/sub, we used it for tons of stuff and the fact it came built in was big as prior to that we had to provision specific stacks to get the same functionality (and it wasn't has nice to use.

memcache has its use but for the little advantages it gave (perf wise) Redis just buried it in developer experience.

Because most of us use Redis for vastly more than a kv. Memcached is excellent if you only have basic cache needs.

redis is a swiss army knife. memcached is a single blade.

I've only used Redis a bit, and memcached not at all yet. But sometimes the superior alternative might lose out simply due to bad luck. I.e. that Redis gained traction and that was it. Or perhaps getting started with Redis is more easy/straight forward than memcached and that's why.

Or it could be any of the features that Redis has which memcached does not that contributed to this outcome.

Try harder ? Noone could explain to you though.

There is a diagram of a process that reads the MySQL commit log and uses that to send Memcache delete commands. Not exactly groundbreaking.

But I would also guess that FB would have to explain lots of their middleware concepts in order to explain how data in the database maps to memcache keys defined in middleware.

Personally, I think the hard part is ensuring the entire company/system accesses data through the correct abstractions (eg. the newest data model code) and does not bypass the data model code (eg. by hitting the database directly).

Not sure about what's been published, but from what I recall, for data stored in memcache it was manual: people would need to know the right keys to invalidate when doing DB mutations (naturally there were functions encoding this, so multiple queries updating, say, a user could all just call a single function which returned the cache keys for that user).

This can be fairly gracefully handled by using a caching library with tag support. On a webpage that collates content from disparate parts of a CMS, you just tag the page content cache with the ID's of all the entities that make it up. Then you define a save/delete handler that purges caches tagged with the ID of whatever you just altered.