Hacker News new | past | comments | ask | show | jobs | submit login
Sparkey – Key/value storage by Spotify (github.com)
109 points by patricjansson on Sept 2, 2013 | hide | past | web | favorite | 55 comments

I remember my interview at Spotify where we discussed how to implement thumbnail display service in the most effective way. What we actually came to is something along the lines of this library.

I always like it when a company focus on their real problems in job interviews and manage to avoid the brain teaser trap. That way you can have a feeling about the job that you are going to work on there, and see if you really like both work and people.

Thanks :-) That is a go-to interview question for us and we actually all do it slightly differently and take it in different directions depending on your expertise or specialization.

It is really closely aligned to what our core service is (distributing and streaming files) and is a great chance to talk with the interviewee and figure out where their strengths are.

Totally encourage other people to interview this way. It's what I've done at the past few companies I've been at and really worked excellently — just think of a problem you're working or have worked on, and distill it into an interview problem.

I also like this style, but you have to be very mindful of the fact that you've been thinking about the problem 1000x more than the candidate. The Curse of Knowledge[1] haunts the interview process. I haven't tried it much, but maybe it would be better to always use a fresh problem in each interview that not even you had seen. Maybe selected from StackOverflow.

1. http://en.m.wikipedia.org/wiki/Curse_of_knowledge

If you're familiar with thumbnail display you know Facebook has Haystack for efficient image serving. It goes very low end. Sparkey or CDB or BAM (mentioned in this post) could be much less complex and can do similar work. Why did they go Haystack route I dont get it.

Are you asking why they didn't use Haystack? Haystack isn't released publicly as far as I can tell (plus, just because another company writes something doesn't mean it's actually good).

No, I think he is asking why Facebook isn't using something like CDB.

Sorry not being clear enough. Yes I was kind of asking why they didnt try simpler options. When you look at their paper on Haystack you see that it's essentially a constant local database.

Another cool project is bam[1], a constant key/value server built on similar principles. A single input file (in this case, a TSV file instead of the SPL log file) and an index file. The cool thing about bam is that it uses the CMPH[2] library to generate a minimal perfect hash function over the keys in the input file before putting them in the index file.

[1]: https://github.com/StefanKarpinski/bam [2]: http://cmph.sourceforge.net/

Wow, didn't expect this to make a mention on the front page of HN today. I never did convince Etsy to let me deploy bam in production, but it's so simple that it should be doable without much fuss. I mainly built it as a proof-of-concept to show that serving static data does not have to be difficult – and that loading large static data sets into a relational database is a truly wasteful, terrible approach. Are you actually using bam "in anger"?

Bam looks really interesting, definitely a lot simpler than Sparkey, and the basic principle is the same. I have been hesitant to use perfect hashing for Sparkey since I wasn't sure how well it holds up for really large data sets (close to a billion keys). Impressive to write it in less than 300 lines of clean code!

There are a bunch of comments related to performance data - the code is available so nothing is stopping anyone from making an unbiased comparison. :)

That said, I intend to publish some sort of performance comparison code / results. The downside with me doing it is that: 1) I know the sparkey code much better than I know level-db or any other solution, so the tuning parameters will probably be suboptimal for the other solutions. 2) I will only focus on our specific usecase (write large bulks, do lots of random reads), which may seem a bit unfair to the most general solutions.

Here are some preliminary performance benchmarks from my regular workstation (Intel(R) Core(TM) i5-2400 CPU @ 3.10GHz, 8 GB RAM): http://pastebin.com/7buZVgdu

The sparkey usage is fairly optimized, but I just randomly put something together for the level-db, so consider the results extremely biased.

Where's the source code for your bench? How large are the records you're loading? How large are the keys? What's the insert order?

How does your test compare to http://symas.com/mdb/microbench/ ? If you're going to try to talk about numbers, talk about them in a meaningful context. Right now you're just handwaving.

Yes, this post was handwaving, and I tried to make that clear ("preliminary", "extremely biased").

On monday I added some slightly more proper benchmark code, you can find it on https://github.com/spotify/sparkey/blob/master/src/bench.c

I didn't add the level-db code to this benchmark however, since I 1) didn't want to manage that dependency 2) didn't know how to write optimized code for usage of it.

I'm using very small records, a couple of bytes of key and value. The insert order is strictly increasing (key_0, key_1, ...), though that doesn't really matter for sparkey since it uses a hash for lookup instead of ordered lists or trees.

As for the symas mdb microbench, I only looked at it briefly but it seems like it's not actually reading the value it's fetching, only doing the lookup of where it actually is, is that correct?

"MDB's zero-memcpy reads mean its read rate is essentially independent of the size of the data items being fetched; it is only affected by the total number of keys in the database."

Doing a lookup and not using the values seems like a very unrealistic usecase.

Here's the part of the benchmark I'm referring to: for (int i = 0; i < reads_; i++) { const int k = rand_.Next() % reads_; key.mv_size = snprintf(ckey, sizeof(ckey), "%016d", k); mdb_cursor_get(cursor, &key, &data, MDB_SET); FinishedSingleOp(); }

It's a valid measurement of how long the DB takes to access a record. You can assume that the time required for an app to use the values will be the same across all DBs therefore it's omitted. All of the microbench code behaves the same on this point.

http://symas.com/mdb/memcache/ and http://symas.com/mdb/hyperdex/ give results where the records are transmitted across the network.

Well, some DB's "load" the value in some way before giving it to the user, so that time is implicitly measured for those types of DB's but not for others, so I don't think it's a particularly fair comparison. I think Tokyo Cabinet gives you a pointer to newly allocated memory, at least for compressed data (but I am not completely sure about this). Like LMDB, Sparkey also does no processing of the value for uncompressed data, but for compressed data some decompression needs to take place in the iterator buffer (I guess that's equivalent to your cursor object). Even worse, if this is done lazily upon value retrieval, the cost is completely hidden in the benchmark.

In any case, I think the easiest way to get a fair benchmark is to at least iterate over the value, possibly also compare it. If that time turns out to be significant (perhaps even dominant) compared to the actual lookup time, then further optimization of the actual storage layer is pretty meaningless.

Have a look https://github.com/hyc/sparkey/tree/master/src bench.c, bench_mdb.c bench.out, bench_mdb.out

This was run on my Dell M4400 laptop, Intel Q9300 2.53GHz quadcore CPU, 8GB RAM. The maximum DB size is around 4GB so this is a purely in-memory test. Your hash lookup is faster than the B+tree, but with compression you lose the advantage.

Thanks, that's interesting data!

I am not sure why you changed the key format to "key_%09d" - is that an optimization for lmdb, to make sure the insertion order is the same as the internal tree ordering? If so, why is that needed for the benchmark?

I noticed that the wall time and cpu time for the sparkey 100M benchmarks were a bit disjoint, it would seem that your OS was evicting many pages or was stalling on disk writes. The Sparkey files were slightly larger than 4 GB while lmdb was slightly smaller, but I am not sure that really explains it on an 8 GB machine.

I am not sure I agree about the non-linear creation time difference, the benchmarks indicate that both sparkey and lmdb are non-linear. The sparkey creation throughput went from 1206357.25 to 1109604.25 (-8.0%) while lmdb's went from 2137678.50 to 2033329.88 (-4.8%)

Regarding the lookup performance "dropping off a cliff", I think that is related to the large difference in wall time vs cpu time, which indicates a lot of page cache misses.

lmdb seems really interesting for large data sets, but I think it's optimized for different use cases. I'd be curious to see how it behaves with more randomized keys and insertion order. I didn't think of doing that in the benchmark since sparkey isn't really affected by it, but it makes sense for when benchmarking a b-tree implementation.

Sparkey is optimized for our use case where we mlock the entire index file to guarantee cache hits, and possibly also mlock the log file, depending on how large it is.

The way you append stuff to sparkey (first fill up a log, then build a hash table as a finalization) is really useful when you need to use lots of memory while building and can't affort random seek file operations, and in the end when most of the work is done and your memory is free again, finalize the database. Of course, you could do the same thing with lmdb, first writing a log and then converting that into a lmdb file.

Thanks for taking the time to adapt the benchmark code to lmdb, it's been very interesting.

Yes, I changed the key format to allow using the MDB_APPEND option for bulk loading. (That's only usable in LMDB for sequential inserts.) Otherwise, for random inserts, things will be much slower. (Again, refer to the microbench to see the huge difference this makes.) If you don't have your data ordered in advance then this comparison is invalid, and we'd have to just refer to the much slower random insert results.

Still don't understand what happened to sparkey at 100M. The same thing happens using snappy, and the compressed filesize is much smaller than LMDB's, so it can't be pagecache exhaustion.

Also suspicious of the actual time measurements. Both of these programs are single-threaded so there's no way the CPU time measurement should be greater than the wall-clock time. I may take a run at using getrusage and gettimeofday instead, these clock_gettime results look flaky.

Could be due to a bug related to reading uninitialized data on the stack. That could lead to using the wrong number of bits for the hash, causing an unnecessarily high number of hash collisions, which makes it more expensive due to false positives that needs to be verified. I think it's fixed in the latest master, and the benchmark code now prints the number of collisions per test case, which could be useful debug data.

Also, I think it would be more interesting to see a comparison with lmdb using random writes instead of sequential.

As for the cpu time measurement, the wallclock is very inprecise, so it could be some small quantum larger than cpu time, but it should never be more than the system specific wall clock quantum.

re: random insert order - if we just revert to the original key format you'll get this: http://www.openldap.org/lists/openldap-devel/200711/msg00002... It becomes a worst-case insert order. If you want to do an actual random order, with a shuffled list so there are no repeats, you'll get something like the September 2012 LMDB microbench results. If you just use rand() and don't account for duplicates you'll get something like the July 2012 LMDB microbench results.

(I've updated my repo using gettimeofday/getrusage).

Other interesting details from the results: LMDB's creation time is always faster. LMDB's creation time is linear, Sparkey's is nonlinear. For a 10x larger DB, Sparkey takes more than 10x longer time to create.

Sparkey's lookup performance drops off a cliff at 100M elements. This doesn't seem to be related to raw size because it occurs regardless of compression. LMDB's performance degrades logarithmically, as expected of an O(logN) algorithm.

Hashing is inherently cache-unfriendly, and hashes are inherently wasteful - hash tables only perform well when they're mostly empty. They're completely hopeless when scaling to large datasets.

and just for curiosity's sake, benchi.c, bench_mdbi.c, benchi.out, bench_mdbi.out using integer keys instead of strings.

And the point of LMDB is that it's zero-copy, there is no wasted time "loading" values.

I was surprised to see LevelDB (https://code.google.com/p/leveldb/) was missing from the list of storage solutions you tried, because it seems optimal for your use-case. Were you aware of it?

I'm not sure about the optimal use-case match. Sparkey is for "mostly static" datasets where on disk structures are generated by a batch process and pushed to servers providing read only access to the data to consumers.

leveldb on the other hand, supports concurrent writes and provides features to handle data consistency and cheap gradual reindexing.

Also Sparkey seems to work well with bittorrent/rsync distribution. I recall spotify use bittorrent to distribute files to their servers.

According to GitHub, last commit (which is also initial checkin) is two years old.

The project is hosted at Google Code. Last commit was Aug 21: https://code.google.com/p/leveldb/source/detail?r=748539c183...

There is also the basho fork of LevelDB (https://github.com/basho/leveldb) which adds a few nice things.

If you're interested in the internals of LevelDB, I strongly recommend watching this talk http://www.youtube.com/watch?v=vo88IdglU_8 (slides here https://speakerdeck.com/basho/optimizing-leveldb-for-perform...)

He means the last commit on sparkey

Right, this is actually the first commit - it's just that the history was squashed before publish - we had to remove sensitive Spotify specific things in it, and it seemed easiest to just do a big squash.

Why do you think it seems optimal for their use case?

A quick glance over LevelDB's features gives me the impression that its bulk-write performance would not be sufficient.

The problem is, that so many projects for so many things exists. It is hard to find the matching project, thats why many people invent the wheele again.

LevelDB has very few good use cases and this certainly isn't one of them. http://symas.com/mdb/microbench/ http://symas.com/mdb/hyperdex/

I wish more projects would follow this kind of readme format, at least somewhat. There are so many new things that popup on HN but have very little information about all the whats and whys I should care

What problem are you solving?

If existing solutions existed what hurdles did you face with them and how did you overcome them with your custom solution?

How do you compare from a performance view? (granted they still need to do this, but at least put in a section about it)

Looks like a cdb variation that moves index to a separate file and therefore allows changing the database (to a degree) without requiring a rebuild.

[0] http://en.wikipedia.org/wiki/Cdb_%28software%29

Indeed, from the README: We used to rely a lot on CDB (which is a really great piece of software). It performed blazingly quick and produces compact files. We only stopped using it when our data started growing close to the 4 GB limit

A 64-bit port of cdb isn't too hard: https://github.com/pcarrier/cdb64

Similar also to DiscoDB, which does support compression, and uses perfect hashing for constant time-lookup with minimal disk access.

Not only that but it provides lightning-fast conjunctive normal form queries, a.k.a logical combinations of primitive keys. Plus it has Python / Erlang bindings.



Yeah, my first two thoughts were discodb and bitcask http://basho.com/hello-bitcask/ too.

I'm baffled by the choice of using the GNU autofools chain just to include a Doxygen target in the Makefile. The whole thing is essentially straight up C with just 1 library dependency.

The command line argument processing is also quite haphazardly done, it's not like it using getopt or whatever that poses compatibility issues. Is writing and packaging with a Makefile that difficult?

We used to have a "simple" makefile, but once we started to support multiple development environments (OSX, various Linux flavors) it got more and more complex. Autotools actually brings a lot of functionality as part of the package and is what people expect. I'm no fan, but it works for our use case.

They wrote a database to solve an operational need. From experience I can tell you that's an endeavour you should strife to spend as little time on as possible.

I think it's a miracle they produced something they feel comfortable sharing with the world. If you write a database in house, and the tool chain and the argument processing are the only things done haphazardly, then hats off to you :)

Whoa I sense passive-aggressiveness :) I'm still waiting for those benchmarks. The code is very clean and simple I was just picking bones. The use case seems to be overly specific tho. Is there any other example use cases where this library could be useful?

Submit a patch. Show them how it could be better.

rm -rf sparkey

apt-get install lmdb

There is LMDB (http://symas.com/mdb/) storage solution for read heavy workloads. Alternative to LevelDB or CDB.

This looks interesting. Can be used for deduplication by keeping this hash table on disk for large amount of data

I wrote something like this to optimize disk seeks heavily by returning a reference of 8 byte and keeping a hashtable in memory. A mostly-append only records store that allowing mutations of same key and by rounding size of blobs by power of 2. Written to optimize storage layer for Membase.


I have now created a very simple benchmark suite to give you some rough performance numbers, and updated the README to include some sample numbers for one specific machine.

I'm struggling to find something that it does that a webserver pointed at the filesystem doesn't do (with hash-ids for file names). I'm wondering if that's all it is, with a bit of logic to write the files in the correct structure.

From the description: Sparkey is an extremely simple persistent key-value store. You could think of it as a read-only hashtable on disk and you wouldn't be far off.

Good point, I missed that. Read the feature list which lists stuff the filesystem does itself.

This goes to prove that simple is fast.

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact