In SQL systems, you can easily shift the computation cost from at read time to at write time. That's what triggers are for. When new data comes in that changes a result you know you need quickly, a trigger can automatically add to a work queue to update a result table.
Since the trigger has automatic access to the contents of the new data (and old data in case of an update or delete) the computation to update the results table can often be made much faster.
Every situation is different of course, but it's overreaching to say that SQL systems have no options beyond read-time computation of results.
First, Cassandra uses a disk layout similar to the one described in the Bigtable paper (http://labs.google.com/papers/bigtable.html sections 5.3 and 5.4); in particular it does no random writes. Relational databases like mysql pretty much all use btree-based storage which was great 20 years ago but is terrible today when seeks are your bottleneck.
I was talking to some people today who are struggling to get mysql to do ~100 insert/update operations per second. Cassandra will easily give you 10x that -- _per node_.
The second benefit is that Cassandra gives you real, scalable partitioning, invisible to the app, for when you do need to add nodes. When you have more than a handful of machines, not having to babysit replication + partitioning is a huge, huge deal.
Because of how the storage engine works, mysql insert performance degrades significantly as table size grows into double digits of GB, and especially once the PK index no longer fits in RAM.
Cassandra doesn't have to read indexes to figure out where to put the data so write performance stays constant as your data set grows.
A 7,200 RPM disk can do about 120 IO operations/s, max. Each transaction requires at least one IO operation to be properly committed to disk.
So ~100 inserts/updates per second is pretty close to the limit of modest hardware, unless you are either 1) doing more than one insert/update per transaction or 2) leaving write caching enabled on you disk.
#1 is the right thing to do when it works for your app, but it doesn't always work.
#2 is the wrong thing to do if you care about the integrity of your data (but its probably not a bad thing to do in development if you understand that you'll probably need a battery-backed caching disk controller to get that kind of performance in production).
Cassandra avoids some of these issues by allowing the developer to specify the desired level of durability. A transaction may have to arrive on disk to be considered complete, or less stringently, it can arrive in memory on one or more additional machines in the cluster.
I'd also guess that cassandra will buffer and batch writes. Postgres can also do this with the write ahead log (WAL). I don't know if mySQL does the same. This trades some (configurable) latency for throughput. A transaction has to wait for the buffer to fill or a timer to expire before it is completed which increases latency, but multiple transactions end up bundled into a single IO operation, which increases throughput.
It's interesting what data structures would be optimal for the SSDs from the near future. If random writes are only a controller problem, it may be the good old btree again.
What's amazing is that 3 terabyes costs about $300 nowadays. Include good components, power, redundancy, maintenance, and it probably comes to $1,000 per year.
So this feature is actually pretty cheap to do, unless I'm missing something.
Every time I read "terabyte", I have to pause, because my brain goes "That's gigantic!" and then a little voice says, "It's the same size as the hard disk in my desktop...it cost $89."
I have the same sort of weird mental disconnect with SD cards. It's just hard to fathom 32GB of data fitting into something about the size of a quarter.
Hey, I know the feeling, it's weird to adjust to but I still haven't filled my 5 Terabytes of space yet so I'm happy. Soon I'm going to offload all my files to a file-server I'm building with an exorbitant amount of HDD space.
Serving up thousands of requests per seconds from 3 TB (while continually adding and updating more data!) is a lot different from just archiving it.
A _lot_. :)
To put it in perspective, 100GB is pretty much the limit of what you can serve from a single mysql machine, for instance. This will vary depending on exact workload but that is a reasonable ball park number.
Of course, in a relational database you wouldn't (generally) be storing the vast majority of the data in question...you'd be querying for it. That's not to trivialize the improvement in performance here, just to point out that this probably would be 100GB in a MySQL database.
"We started thinking seriously about deploying Cassandra in production around three weeks ago. After looking at the site for something that would be a good fit, we settled on green badges."
It seems completely baffling to me that someone would go out and compare different db solutions, pick one and THEN try to find a way to fit it on their site architecture.
I read that as, "which feature is easiest to port from the existing mysql solution?" Which is a reasonable question to ask, given how different the two databases are.
SELECT `digdate`, `id` FROM `Diggs`
WHERE `userid` IN (59, 9006, 15989, 16045, 29183,
30220, 62511, 75212, 79006)
AND itemid = 13084479 ORDER BY `digdate` DESC, `id` DESC LIMIT 4;
Ok, how do we optimize this query?
Step 1: Keep all dugg items in memcached for the last N days, where N is determined by when you run out of memory. Then, your query becomes:
SELECT `digdate`, `id` FROM `Diggs`
WHERE `userid` IN (59, 9006, 15989, 16045, 29183,
30220, 62511, 75212, 79006)
AND digdate < now() - interval '5 days'
AND itemid = 13084479 ORDER BY `digdate` DESC, `id` DESC LIMIT 4; /* Excuse the postgresql syntax /
If your database is properly clustered, this will mean you are only running the query against partitions holding old dugs, which is probably not as hot as the more recent stuff. Additionally, I strongly suspect that you see more recent articles more than old ones, if the article is less than 5 days old you need no SQL at all, just the memcache lookup. For example, if you are looking at the homepage, and there are 15 articles on it, you have to do a single memcached get request for all the pairs like (article_id, friend_id), so if you have 100 friends that is 100 15 keys to request. This is large, but who cares, you can add memcached servers and webservers until you puke and this will keep scaling without limit. When browsing old articles the db will get hit heavily, but only the partitions holding old data, and I would guess that this is a very very small fraction of their overall use.
Step 2: When a user is activly using the site, like they have viewed 2 pages in the last 10 minutes or something, shove all their old (article_id, friend_id) pairs into memcached as well. Once a user has reached the 'activity threshhold' and the cache is filled, no sql is necessary to find all their friend's dug articles. As a bonus, no weirdo software like 'cassandra' which may or may not continue to exist in 1 year is necessary.
For step 1 you need very little effort, just put a key into memcached every time a user digs something, and put a 5 day timeout on that key. This is 1 line of code in whatever code handles the http request representing a 'dig'. Then you have to build up the list of friends and keep it somewhere when a user logs in to the site (or returns with a cookie that has them logged in). This would take one memcache request when the user logs in/comes back to see if their friends list is in memcached, one sql statement if it is not, and a line in the are that handles adding friends to spoil the key if their friends list changes (you could try updating it, but why, just let it be regenerated on their next http request). Finally, you have to generate the keys for the (article_id, friend_id) pairs on each page view, and do a multi_get from memcached.
Step 2 would require an asynchronous process, so would be more complex.
I could implement step 1 in an hour or so if familiar with the digg codebase, and step 2 in perhaps 2 days, however if they have other async processes that occur when a user logs in that you could integrate this with it could take as little as an hour or two as well, since the logic is dead simple, it is the mechanics of running a process to do it that is time consuming.
Finally, you would have to figure out how much memory you would need to store N days of digs (users with no friends do not count in this). I believe it would not be very much.
Since I wrote that query back in 2005, and it was actually somewhat more complex (had to deal with users banned after they dug the same story as you for starters) I assume Ian (whom I've never met) is simplifying for pedagogical reasons. And we had something similar to what you describe in no. 2 in 2006.
Also I find it fascinating that the table is "Diggs" (it was originally "digs" and I suspect it still is). Gotta get the branding right, even in sql queries.
When a Memcached instance goes under, all Diggs stored on that instance in the last N days disappear. Not the end of the world for this application, but very undesirable for most things. You could potentially "fall back"
to MySQL for this but a workable strategy for that probably isn't simple.
Further, step two is irrelevant. There's no way of knowing if a (article_id, friend_id) pair in memcached is from the most recent N days or whether it's been stuffed in due to a user being active. Therefore, searching the DB for older diggs is still necessary, and should take the exact same amount of load as if they weren't in memcached at all.
Memcached + MySQL makes great sense when the data set is small and simple. If all the content on the site fits within 1 GB you could probably easily push a hundred millions unique visitors a day. For an application like this, the relatively poor performance of MySQL and the inflexibility of Memcached cause problems.
It seems to me that using Cassandra, even in its current immature state, makes much more sense than the solution you're proposing.
Ahh, you make a great point, that if an instance goes down you lose those keys. Cassandra doesn't necessarily solve this, however, unless you have enough failover that it is unthinkable that you could have cassandra lose data. If you can rebuild the cache in cassandra you can rebuild it in the identical way using my scheme, but my scheme only requires N days of data to be run through.
Step 2 is fine, memcached allows setting of timeouts on keys. If you always set a timeout of (dig_time + five_days) - now() you are set. Be careful, times > 30 days imply a unix timestamp not a timeout.
Mysql kindof sucks. The clustering is easy to set up at least.
Cassandra and other big on disk hash tables are pretty cool. I think once they have datastore-like indexing capabilities they will be totally usable. My qualm with using them is that there are 80 of them right now, and they are all pretty immature. The ones that are mature, like bsddb, and complicated to use.
My point was that cassandra is just a big memcached.
> unless you have enough failover that it is unthinkable that you could have cassandra lose data
That's pretty much the idea. Cassandra makes replication + failover totally seamless, so there's no excuses in that respect. :) Cassandra also supports replicating across multiple data centers.
NoSQL is a buzzword, originally to describe "distributed open source databases" but it's catchy enough that pretty much every nonrelational db uses it.
So saying "NoSQL gives you [or doesn't give you] X" is meaningless.
And for the record, SQL is great, but trying to keep using it once your data set has outgrown a single machine is like "trying to build a bookshelf out of mashed potatoes." (Gotta love jwz.)
Anyone from digg here? I'd love to know how long it took to build that dataset the first time. In other words: what's the recovery window like to reset that size dataset?
Why does the first step "Query Friends for all my friends" take 1.5 seconds? I am struggling to understand this. If this simple table of say 100,000,000 rows is indexed on userid and we are only reading and returning say 200 rows for a particular userid what makes it so slow?
But surely that's broken. If I dig something but then add a friend AFTERWARDS, they wont see the shield as the bucket for my digg didn't contain their user id at write time. Am I missing something?
Since the trigger has automatic access to the contents of the new data (and old data in case of an update or delete) the computation to update the results table can often be made much faster.
Every situation is different of course, but it's overreaching to say that SQL systems have no options beyond read-time computation of results.