Adventures in Container Sharding – SQLite performance problem and the pivot point.

Hey world it’s been a while, turns out I’m not much of a blogger. But I know how useful for myself it is to do write-ups occasionally so I can actually find them later.. having said that. In my last post I mentioned I was an OpenStack Developer.. and it’s still true. I spend my time hacking and working on Openstack Swift the awesome OpenSource object storage cluster.

One thing I’ve been trying to tackle recently is container sharding in Swift, I will not go into full details as there is a Swift Spec that is relatively recent, and I’ve also gave a high level talk on it at LCA in Geelong.

The tl;dr being, Swift accounts and containers (or the metadata layer of Swift) are SQLite databases that get treated like objects themselves and replicated throughout the cluster. Which works amazingly well. Until you add millions and millions of objects to a container. And what I’m talking about here is container level object metadata, not the objects themselves. When this happens, SQLite being a file starts to have latency and locking issues, as one would expect. The solution to this is shard up these container databases throughout the cluster, which is what I’ve been working on.

At the last OpenStack summit in Austin, the awesome people at SwiftStack, whom I work quite closely with in the community gave me a container database they generated that has 700,000,000 objects in it (metadata again). This SQLite file is about 105G so not small. Plugging this into a small cluster I have to test my sharding implementation has been interesting to say the least.

When sharding a container down, we have a simple idea, split it in half. That is to say find someplace in the object table to pivot on. We can then keep pivoting giving us a list of ranges (which can be treated as a binary tree). The problem is finding the pivot point. In all my testing up til now I had what I thought was the perfect and simple way:

FROM object
WHERE deleted=0 ORDER BY name LIMIT 1 OFFSET (
SELECT object_count / 2
FROM policy_stat);

This did amazingly well in all my tests.. but I obviously never got big enough. This simple SQL statement would do plenty well if sharding in Swift was turned on from day dot, but the sharding plans for Swift is for once it’s solved in this POC to add to Swift as a beta which can be turned ‘on’ at a container by container basis when you want. After it graduates from beta but is still a switch. To finally once we are confident in it’s ability to have it on permanently. In the latter case container’s would never get big enough to worry about.. However, in the earlier stages a user would only turn it on when the container is _very_ slow.

Using the pivot SQL statement on the large container I was given ground to a halt, I’m sure it would have come back to be eventually, but I got tired of waiting after what seemed for ages.. there has to be a better way.

Turns out the OFFSET statement in SQLite, even when hitting an index still does a scan to find the offset.. This is slow when you get to a very large table size. Turns out under the hood, the resultset is stored as a double-linked list, and OFFSET will still scan down the results, which I’m sure probably has optimisations but anyway I was struggling to think of a good way to find a good enough middle value that didn’t involve some table scanning. You can see from the SQL statement, we know have many objects we have in the container, but the problem is because swift is eventually consistent we need to temporally store objects that have been deleted. So randomly picking an index doesn’t help, and it wont necessarily be in name order.

So on really large containers OFFSET needs to be thrown out the window. Turns out the sharding implementation can deal with shrinking the number of shards, merging smaller ranges together, not just growing/splitting. This means we don’t actually need to be exact, we also don’t actually need to split on an existing object, just a name that would be somewhere in the middle and so long as it’s cutting down the large container then it’s good enough. So what can we do?

Turns out there is an optimisation in SQLite, because an index is a double-linked list and ordered by it’s index, it’s really quick if all we want to do is go to the first or last element. So that’s what I’ve done:

SELECT min(name) as name FROM object WHERE deleted = 0;
SELECT max(name) as name FROM object WHERE deleted = 0;

These two statements are blindingly fast due to the fact that we already have a compound index on name and delete (for cleaning up). Note however they have to be run as 2 separate commands, combine the two into one and you loose your optimisation and you’ll have to scan all elements. Having the min and max name is a good start, and even when dealing with already sharded containers, they are just smaller ranges so this still works. The question is now what?

In the perfect work we have an even distribution of objects between the min and max names, so we just need to find a middle name between the two to pivot on. Turns out even in a not evenly distributed container we will still be shrinking the container, even at worst case only be a few objects. But these will be cleaned up later (merged into a neighbour range by the implementation). And so long as the container gets smaller, eventually it’ll shrink small enough to be usable.

Next step is finding the middle value, to do this I just wrote some python:

from itertools import izip_longest
import sys

lower = unicode(sys.argv[1])
upper = unicode(sys.argv[2])

def middle_str(str1, str2):
    result = []
    for l, u in izip_longest(map(ord, str1), map(ord, str2), fillvalue=0):
        result.append((l + u) // 2)
    return u''.join(map(unichr, result))

if __name__ == "__main__":
    print(middle_str(lower, upper))

What does it do. Calling middle_str(min, max) will grab the unicode versions of the strings, turn them into there interger values, find the middle and turn them back into a word. After matching the prefix that is. So:

$ python 'aaaaaaaaa' 'zzzzzzzzz'

$ python 'aaaaaaaaa' 'aazzzzzzz'

$ python 'DFasjiojsaoi' 'ZZsdkmfi084f'

I am now plugging this into my implementation and lets tackle this large container again.