Yeah I have learned about some people have different perceptions of what cache invalidation means (I have a narrower definition, a subset of what some people think as cache invalidation). I will not die on this hill. I am happy to rename/rephrase/anything to be helpful.
> it doesnt sound like Polaris would handle a network partition well ...
Good question. And you are right. My thought on this is that over time, polaris would still catch those unique issues happened at network partition or whatnot. Polaris doesn't promise have 100% coverage all the time. However, over time, if there are flaws in the system, Polaris should help surface it. And once it did (even with just one inconsistent sample), it should be very actionable, and we can find out why it happened, and fix the underlying cause.
> I also feel like the consistency tracing is assuming that only one fill and invalidate is occurring at a time (when in practice, there may be multiple fills and invalidates occurring in parallel on the same data)... and that those calls will arrive in order. If they arrive out of order, it doesnt sound like it would catch that..
It doesn't make that assumption actually. Distributed systems are state machines. Consistency tracing essentially logs (it doesn't do much detection) state transitions. So when polaris detects an anomaly, we have all the information to help us diagnose. And you are right that fills and invalidations can happen in parallel and it's fine. E.g. if fill happens before invalidation, we can always track state mutations caused by the invalidation. If invalidation is the culprit (the earlier fill can't be because it happened earlier), we would have a log for it.
Sorry if it sounds like silly questions, I was genuinely curious to learn about the strategy since the blog started with "When it comes to cache invalidation, we believe we now have an effective solution to bridge the gap between theory and practice"
> what about many-to-many where one fetch returns inconsistent results Are you referring to the case when cache stores a complicated result of a function that fetches data from many sources, and how to figure out when to invalidate? If so, I am sorry but that's not the cache invalidation I was referring to. (https://news.ycombinator.com/item?id=31676102)
> What about data queried while there's a transaction rollback.. when is it updated in the cache
This might be easier to answer. I assume if you run the database with isolation level higher than read-committed, you will be fine.
Could this delay cause a wrong reporting? Since the source of truth (the DB) could already received multiple mutate events of the data Polaris trying to verify? How do you handle this case?
Assuming your definition of cache invalidation is about "when/who" to invalidate on writes. Let's actually try solving it.
E.g. in its most generic form, a cache can store arbitrary materialization from any data source. Now when updating the data source, in order to keep caches consistent, you essentially need to transact (cross system transaction) on both the data source and cache(s). Usually cache has more number of replicas, I am not sure running this type of transactions is practical at scale.
What happens if we don't transact on both systems (the data source, and cache)? Well, now whenever the asynchronous update pipeline performs the computation, it's done against a moving data source (not a snapshot of when the write was committed). Now let's say the data source is Spanner, which provides point-in-time snapshots. On Spanner commit you can get a commit time (TrueTime) back. Now using that commit time, to read the data and compute cache update asynchronously can be done.
Because the materialization in cache don't take writes themselves, so updates to them are essentially performed blindly and can be ordered by the commit time (TrueTime).
Now this does assume whatever we cache (the query e.g.) needs to be schematized, and made known to the invalidation pipeline. I think it's a very fair assumption to make. As otherwise (anyone can cache anything without the invalidation pipeline knowing at all), it's pretty obvious that this problem can't be solved.
- client starts a transaction
- client runs any mutations as needed
- client collects user_ids whose cache entries need to be invalidated
- client invalidates cache
- client commits the transaction
You can now have caches fetching & storing the "old" state between the last two steps. This fails to invalidate cache in all scenarios.You can use a versioning scheme supported by the database to do compare-and-swap – i.e. sending invalidate along with a hybrid-logical-clock. HLC is nice in this case as it handles DB transaction rollback gracefully, if the database supports it.
Or we can always do the invalidation asynchronously (by recording the invalidation keys transactionally and have a tailer that sends out the invalidation).
I will make an edit to the blog to make it more clear. Let me know if that makes sense!
I mean this – cache invalidate/fill race specifically – is very much a solved problem from a protocol perspective, as long as our definition of the problem is the same – do not leave stale data in cache indefinitely. That is not to say it is easy; I am not trying to minimize anyone's struggles. https://research.facebook.com/publications/scaling-memcache-... might be of interest.
A lot of the challenges in cache are in making the tradeoffs between consistency and coordination overhead based on the workload and the requirement, _and actually_ making caches consistent in production. As explained in the post, there are practical challenges that are very unique to cache and cache invalidation.
Can't we assume they tested it before deploying to production by introducing all the weird stuff that might happen in a distributed system like partitions, high latency, so they aren't relying only on "it seems to work in production"?
Having better tests and better observability in production are also not mutually exclusive.
And regarding those tools:
it doesnt sound like Polaris would handle a network partition well. If a cache invalidation triggers it to check the other caches for consistency... that assumes Polaris will receive that invalidation message. Imagine a scenario of 5 cache servers, and a Polaris server. On one side of the split is Polaris and 2 servers, and the other has 3 cache servers... It's possible for the 3 cache servers to receive an update that is not received by the polaris+2 network... And polaris would not just be unaware of the inconsistency, but it also wouldn't know to check later for the inconsistency when the network partition is resolved.
I also feel like the consistency tracing is assuming that only one fill and invalidate is occurring at a time (when in practice, there may be multiple fills and invalidates occurring in parallel on the same data)... and that those calls will arrive in order. If they arrive out of order, it doesnt sound like it would catch that.. and I think you're relying on Polaris to catch this case, but high latency cannot be differentiated from a network partition except the length of the delay... so these two types of errors can be seen together.. in which case, you'd have a cache error that neither tool would detect.
I would like to hear why Im wrong.
I understand this is being used in prod, but network partitions don't occur everyday... and Im not convinced this has seen enough flaky networks to work out the bugs.