Groupcache: an alternative to memcached, written in Go
github.com
github.com
- thundering herds suck, so it manages how the cache is filled when a key is missing
- repeatedly retrieving super hot values will eventually exhaust the network pipe of whichever server held, so groupcache replicates hot values to more than one cache server
Since this was built for dl.google.com, the 'value' might be big (say, a chunk of the Chrome binary) and the request rates for the top cache keys probably get pretty high. So, even if memcached is doing good work for lots of folks, the distinct features of groupcache probably address real needs in the app.
And, yeah, groupcache and the dl.google.com rewrite are Go partly because Brad likes Go, but both projects had reasons to exist quite apart from deploying Go for Go's sake.
I also then ported the C++ memcached (called memcacheg) to Go, for profiling Go vs C++, and named that one memcachego. So I've written it about 4 times.
Kidding aside, anything you learned during the rewrites that caused noticeable changes to the design, or it's still the same thing, just transliterated?
How were the results of the profiling? Anything interesting come out?
The others sound like they are mostly used by Google.
(* ) http://cdn.parleys.com/p/5148922a0364bc17fc56c60f/GarbageCol...
If you're worried about this, stick with C++.
The situation you describe, as you describe it, is (no longer - was it ever?) a pathological case. I run stuff with 12-20GB active objects and a few dozen http hits a second and the only times I get a latency over 100ms is when I reload/refresh some data from disk (1.5GB of gob), then it goes to 150-250ms typically.
It does not matter how sophisticated GC code is when you have millions of tiny objects from from crappy Java code with useless copying and referencing. This is why it uses so much memory.
It is very difficult to find any worse thing that Java in terms of wasting of memory. Take any Hadoop none and measure the ratio between memory used by Java processes and amount of actual data stored in memory. Near factor of 2?
Or, you know, you can avoid writing "crappy" code. Whereas with Go's GC, even proper code will have problems.
You are arguing something that is beyond discussion: Java's GC is better than Go's.
Even in Clojure code, for example, they implement a node of a binary tree as a hash-table which is a meaningless waste. Position based selectors would be good enough.
Java's GC is better than Go's - yes, in theory, on paper. In practice well-written Go code could outperform typical Java code, even with less sophisticated GC.
Of course, I cannot prove that, but there are some intuitions to support such claims.
It is true that you should describe Go as a garbage-collected language, but unlike a lot of other such languages, Go has C-like mutable arrays readily available. In practice it's more like a sort of hybrid, in that even though it's garbage collected you have a lot of opportunities to write code that still doesn't really use it, without having to "drop down to" C or something. I still hope to see some improvements in it before I could make a big commitment to it, but I've certainly got my eye on it.
Nothing stops you, except: - The determination NOT to do something the computer can do - Previous exposure to a proper, modern GC - The fact that you used Go to get away from this kind of shit in the first place
If you don't want to do that, don't. I'd never start the prototype with that functionality, for instance. But when you discover that you need it, Go permits it in a way that Python or Perl do not. Go is GC'ed and you are free to use that to the extent you want, but it's easier to escape from the GC than it is in many GC'ed languages. It may not be obvious from a casual reading of the spec, but there's a lot of ways around garbage in Go. Not like, say, Rust, and certainly not the same level of Raw Unstoppable Power as C++ (at a corresponding huge complexity cost), but it's a very interesting middle ground.
Note that if you were writing groupcache in Java, you'd probably end up doing the same thing and just allocating an expanse of bytes which you manage yourself.
I've often heard people say things that imply that "proper" GC's don't behave like whatever badness I'm experiencing currently. Can you give me an example of a Garbage collector that I could safely use for a large heap of small cached (ie. long lifetimes) objects without a significant throughput penalty and with deterministic latencies (ie., no random full system pauses) over, say, a tcmalloc based hand-allocated/free'd system ?
I'm quite interested in the two mitigations you mentioned. Especially the "manage memory on your own option" ? I was not aware that there was a go language (that is, not a C extension) blessed way to do such a thing.
As for "reduce/avoid" creating garbage, how can I achieve that for a large heap caching application (like the one we're talking about here) ?
Some careful thoughts about the memory layout of your structs, especially which of them should be embedded and / or passed around by pointers and which of them shouldn't, might also pay off.
Another common optimization is to put the allocated objects back to a memory pool for later use. Take a look at the bufCache channel [1] from the bufio package for example (the http and the json package are using the same trick).
Another sometimes-useful way to save allocations--not for a cache, but in general--is just to turn short-lived allocations into longer-lived ones: if fooBuf is used and thrown away by each of several calls to obj.baz(), then make a single fooBuf and store it as a (private) field on obj, assuming that doesn't present thread-safety or other problems in the specific context.
There are certainly things you wouldn't use Go for, but now you know more about reducing memory pressure in GC'd languages. :)
The problem is, none of those things help when you have a large cache of small(ish) objects that need to be scanned fully at every GC cycle. There's nothing worse for your performance than stopping the world for a few seconds and using that time to wipe your L1/2/3 caches with completely useless data. That's the reason I called large-heaped caches the anti-pattern for a GC'd system.
Brad's answer about using large byte array allocations (and presumably, managing the smaller chunks manually) is a fair enough answer to this question. But of course in this case you'll be writing code to manage small allocations out of that big buffer yourself, thus negating the whole point of GC. But it might be a decent enough compromise if you like the rest of the language a lot (which Brad clearly does :-)
For dl.google.com, the first user of groupcache, it sounds like the Go runtime was working with relatively few, relatively large chunks of memory (think 2MB chunks of a Chrome binary). Since that doesn't have so many pointer-containing objects, just big chunks of inert data that don't need scanned, it's easyish. On the other hand, something like throwaway2424's case could be a GC stress test--many gigs of RAM holding a network of small, pointer-filled objects.
Really, the tl;dr may be to prototype/measure if you need to know how a GC will do. Assumptions are easy; data is hard.
And as for measuring... it's not cheap (developer time wise) to build a realistic enough model of your application and put realistic enough loads on it for long periods of time to test out each of your hypotheses (and GC theories do require long periods of load). So you have to narrow the space of choices informed by past experiences and sometimes by gathering semi-reliable folk lore. I'm currently engaging in the latter activity :-)
Atom also says 1.0's was more conservative, but, as Brad also said, still didn't scan "objects such as []byte" (meaning all plain-old-data arrays? who knows). The Go 1.1 Release Notes mention the collector becoming more precise, which was a particular issue on 32-bit because big heaps could span a lot of the address space.
You can see the GC source itself doing some per-type switching: https://code.google.com/p/go/source/browse/src/pkg/runtime/m...
At some point, this sort of discussion probably gets you less useful info per unit effort than just playing with a Go distribution, trying out whatever toy programs you find interesting.
And yes, unfair to ask someone to test every guess they have. But if you do want to know a bit more about what effect GC and Go would have, experimenting with toy programs isn't a bad way to get a feel.
Parts of your application may rely on the expiration feature. But the biggest change is the inability to overwrite a current cache key. Every application I've used does this constantly (object updates).
Groupcache in its current form is useful for a very narrow set of applications.
I'm writing a torrent tracker in Go that could make great use of a distributed cache. This would enable you to put an arbitrary number of "tracker nodes" behind a load-balancer.
One of the goals of my project is that its easy to configure: everything except the RDBMS is hosted inside a single binary and configured with a single file. In the simplest case: that binary is responsible for two web-servers, a process-local cache, and a DB connection pool. (In more complex cases, each binary simply acts as another node behind a load-balancer; and they attempt to cooperate together.)
It's useful to avoid hitting the disk by having the tracker cache torrents that are active. When you start adding more tracker nodes, this is where having a distributed cache would be nice.
Without groupcache, that means I've just forced my users to install and configure memcached. With groupcache, the configuration and server are simply part of my application. (As an added bonus, since Go is statically linked, I haven't even added any external dependencies on shared libraries.)
---
I agree that the use-cases for groupcache are far more limited in scope, but I was still glad to hear that this was a library, and not a standalone server. I think we need _more_ distributed caches implemented as libraries, not less.
There are many use-cases where it may not be preferable to have a separate caching server. In my career: getting approval to install memcached on a customer's server could add _weeks_ to our deploy time simply because of red-tape; in my personal programming endeavors: setting up memcached is just an added layer of complexity that could easily be avoided by using libraries instead of servers.
I'd argue that we should leave memcached [more specifically: standalone distributed caches] for the more complicated scenarios where a dedicated solution is called for. Bring on the distributed-cache libraries where a much simpler, ad-hoc solution would be more beneficial to end-users.
The only bit which could be tricky is if you want to clear the cache, I don't see a way of doing that from a quick look at the API. But to clear the cache you could have a special version key hashed into each key whose sole purpose is to empty the cache, which you bump whenever you want to remove all keys at once for whatever reason (in a perfect world, this shouldn't be necessary of course).
Given that there's an 'lru' folder with a package that says 'implements a LRU cache' as the first comment, I assume this works that way as well.
An example of a good transaction ID would be SHA1(data).
In general, how can one learn to think in an immutable fashion to effectively exploit this?
For example git has loads of immutable data (every commit, tree and blob is immutable), and only very little mutable data (the refs) that point to some of those immutable objects.
You simply wouldn't use groupcache for session data; you'd use memcached.
Mind you, doing so avoids the hardest parts of caching (and especially distributed caching, which otherwise begins to underperform around ≥ 5-7 nodes), so I can see significant upside. No surprise stales, distribution update clogging, etc.
Its an alternative to memcache but not a direct replacement. I hope he adds CAS etc.
I hope they start using the kernel's buffer cache as the backing store, or explain why its not a good idea: http://williamedwardscoder.tumblr.com/post/13363076806/buffc...
Its nice if you can use groupcache as a memcache replacement even if some use-cases tie the hands of the implementation e.g. replication of hot items.
The CAS must have an authoritative node (my mind wanders thinking about replication and failover) but the key it protects - with the version baked in - can be replicated surely?
I think the failure mode of a system-level cache is rather better than per-application islands that can conflict.
If program A hits swap, it means that cold pages are written to swap so that A can get those pages; this initial writing is done by program A, its true. But A may not be the cause of the problem, A is just the straw that breaks the camel's back.
And those pages that got written to swap likely belong to others, and they pay the cost when they need those pages back...
In my practical experience, when one of my apps hits swap, the whole system becomes distressed. It is not isolated to the 'offender'.
You can of course avoid swap, but with your OS doing overcommit on memory allocations, you are just inviting a completely different way of failing and that too is hard to manage. You end up having to know a lot about your deployment environment and ring-fence memory between components and manage their budgets. If you want to have both app code and cache on the same node - and that's a central tenet of groupcache - then you have to make sure everything is under-dimensioned because the needs of one cannot steal from the other; your cache isn't adaptive.
That's why I built a system to do caching centrally at the OS level.
I hope someone like Brad is browsing here and can make some kind of piecing observation I've missed.
> 64 MB max per-node memory usage
So this is best used as a LRU cache of hot items.
It doesn't compete/replace memcache comprehensively, but it does attack the use of memcache as a relief for hot items.
I can see me mixing my Go programs with both groupcache and memcache.
Edit: I have glanced through the code and cannot see where the 64 MB per-node limit comes in. Anyone see that?
In that case... for me it fully replaces memcache for how I'm using memcache today.
Uh oh. The dreaded cast we see here http://how-bazaar.blogspot.co.nz/2013/07/stunned-by-go.html
Is this anything like how vector clocks are used, where the client uses the clocks to figure out which is the right state in a distributed system?
* it deadlocks. Often.
* it often doesn't recover from network partitions. Symptoms include both sides of the partition never merging, and deadlocks. Recovering requires rebooting the entire cluster at once. Due to their locking strategy, a rolling reboot of the cluster isn't enough.
*some of the developers didn't seem to demonstrate understanding of races, or distributed systems. When I reported racing bugs, they asked if I could attach reliable unit tests.
Has anyone used both/either?
Values stored are immutable, and expire when they are no longer used. There is no concept of a TTL.
> How do you set the maximum size of the process's cache?
The cache is broken into a set of named groups, each with their own separate cache and cache-filling method. The size of each group's cache is set when created via NewGroup[1].
--