Configuring replication (by )

Storing all your data on one disk, or even inside one computer, is a risky thing to do. Anything stored in only one, small, physical location is all too easily destroyed by flood, fire, idiots, or deliberate action; and any one electronic device is prone to failure, as its continued functioning depends on the functioning of many tiny components that are not very easily replaced.

So it's sensible to store multiple copies, ideally in physically remote locations.

One way of doing this is by taking backups; this involves taking a copy of the data and putting it into a special storage system, such as compressed files on another disk, magnetic tape, a Ugarit vault, etc.

If the original data is lost, the backed-up data can't generally be used as-is, but has to be restored from the backup storage.

Another way is by replicating the data, which means storing multiple, equivalent, copies. Any of those copies can then be used to read the data, which is useful - there's no special restore process to get the data back, and if you have lots of requests to read the data, you can service those requests from your nearest copy of it (reducing delays and long-distance communication costs). Or you can spread the read workload across multiple copies in order to increase your total throughput.

Replication provides a better quality of service, but it has a downside; as all the copies are equally important, you can't use cheaper, slower, more compact storage methods for your extra copies, as you can with backups onto slower disks or tapes.

And then there's hybrid systems, perhaps were you have a primary copy and replicate onto slower disks as a "backup", while only using the primary copy for day-to-day use; if it fails then you switch to the slower "backup replica", and tolerate slower service until a new primary copy is made.

Traditionally, replicated storage systems such as HDFS require the administrator to specify a "replication factor", either system-wide or on a per-file basis. This is the number of replicas that must be made of the file. Two is the minimum to actually get any replication, but three is popular - if one replica is lost, then you still have two replicas to keep you going while you rebuild the missing replica, meaning you have to be unlucky and have two failures in quick succession before you're down to a single copy of anything.

However, this is a crude and nasty way of controlling replication. Needless to say, I've been considering how to configure replication of blocks within a Ugarit vault, and have designed a much fancier way.

For Ugarit replication, I want to cobble together all sorts of disks to make one large vault. I want to replicate data between disks to protect me against disk failures, and to make it possible to grow the vault by adding more disks, rather than having to transfer a single monolithic vault onto a larger disk when it gets full.

But as I'm a cheapskate, I'll be dealing with disks of varying reliability, capacity, and performance. So how do I control replication in such a complex, heterogeneous, environment?

What I've decided is to give each "shard" of the vault four configurable parameters.

The most interesting one is the "trust". This is a percentage. For a block to be considered sufficiently replicated, then copies of it must exist on enough shards that the sum of the trusts of the shards is more than or equal to 100%.

So a simple system with identical disks, where I want to replicate everything three times, can be had by giving each disk a trust of 34%; any three of them will sum to 102%, so every block will be copied three times.

But disks I trust less could be given a trust of 20%, requiring five copies if a block is stored only on such disks - or some combination of good and less-good disks.

That allows for simple homogeneous configurations, as well as complex heterogeneous ones, with a simple and intuitive configuration parameter. Nice!

The second is "write weighting". This is a dimensionless number, which defaults to 1 (it's not compulsory to specify it). Basically, when the system is given a block to store, it will pick shards at random until it has enough to meet the trust limit of 100%. But the write weighting is used as a weighting when making that random choice - a shard with a write weightinh of 2 will get twice as many blocks written to it as a normal block, on average.

So if I have two disks, one of which has 2TiB free and the other of which has 1TiB free, I can give a write weighting of 2 to the first one, and they'll fill so that they're both full at about the same time.

Of course, if I have disks that are now completely full in my vault, I can set their write weighting to 0 and they'll never be picked for writing new blocks to. They'll still be available for reading all the blocks they already have. If I left the write weighting untouched everything would still work, as the write requests failing would cause another shard to be picked for the write, but setting the weighting to 0 would speed things up by stopping the system from trying the write in the first place.

The third parameter is a read priority, which is also optional and defaults to 1. When a block must be read, the list of shards it's replicated on is looked up, and a shard picked in read priority order. If there are multiple shards with the same read priority, then one is picked at random. If the read fails, we repeat the process (excluding already-tried shards), so the read priority can be used to make sure we consult a fast, nearby, cheap-to-access local disk before trying to use a remote shard, for instance.

By default, all shards have the same read priority, so read requests will be randomly spread across them, sharing the load.

Finally, we have a read weighting, which defaults to 1. When we randomly pick a shard to read from, out of a set of alternatives with the same priority, we weight the random choice with this weighting. So if we have a disk that's twice as fast as another, we can give it twice the weighting, and on a busy system it'll get twice as many reads as the other, spreading the load fairly.

I like this approach, since it can be dumbed down to giving defaults for everything - 33% trust (for a three-way replication), and all the weightings and priorities at 1 (to spread everything evenly).

Or you can fine-tune it based on details of your available storage shards.

Or you can use extreme values for various special cases.

Got a "memcached backend" that offers fast storage, but will forget things? Give it a 0% trust and a high write weighting, so everything gets written there, but also gets properly replicated to stable storage; and give it a high read priority, so it gets checked first. Et voila, it's working as a cache.

Got 100% reliable storage shards, and just want to "stripe" them together to create a single, larger, one? Give them 100% trust, so every block is only written to one, but use read/write weightings to distribute load between them.

Got a read-only shard, perhaps due to its disk being full, or because you've explicitly copied it onto some protected read-only media (eg, optical) for security reasons? Just set the write weighting to 0, and it'll be there for reading.

Got some crazy combination of the above? Go for it!

Also, systems such as HDFS let you specify the replication factor on a per-file basis, requiring more replication for more important files (increasing the number of shard failures required to totally lose them) and to make them more widely avilable in the cluster (increasing the total read throughput available on that file, useful for small-but-widely-required files such as configuration or reference data). We can do that to! By default, every block written needs to be replicated enough to attain 100% trust - but this could be overriden on a per-block basis. Indeed, you could store a block on every shard by setting a trust target of "infinity"; normally, when given a trust target it can't meet (even with every shard), the system would do its best and emit a warning that the system is in danger, but a trust target of "infinity" should probably suppress that warning as it can be taken to mean "every shard".

The trust target of a block should be stored along with it, because the system needs to be able to check that blocks are still sufficiently replicated when shards are removed (or lost), and replicate them to new shards until every block has met its trust target again.

Tell me what you think. I designed this for Ugarit's replicated storage backend and WOLFRAM replicated storage in ARGON, but I think it could be a useful replication control framework in other projects, too.

The only extension I'm considering is having a write priority as well as a write weighting, just as we do with reads - because that would be a better way of enforcing all writes go to a "fast local cache" backend than just giving it a weighting of 99999999 or something, but I'm not sure it's necessary and four numbers is already a lot. What do you think?


  • By John Cowan, Sun 26th Oct 2014 @ 4:51 pm

    This is brilliant.

    I assume "set the write priority to 0" should be "set the write weighting to 0".

    An actual write priority ought to indicate which of the disks that are going to be written will be written first. As such, it can be inferred from read priority and trust, I think.

  • By Alaric Snell-Pym, Thu 13th Aug 2015 @ 10:30 am

    An IRC discussion (with zaniyah in #chicken on Freenode) raised the issue of support for hierarchic storage - fast SSD, fast disk, slow disk, tape - which the above model doesn't quite cover, as it doesn't handle the migration of blocks based on access frequency. This is also not quite handled properly in the "memcached backend" example above, which fails to cache blocks when they are read from other shards...

    I need to think about this! Do we need another shard parameter, or will the existing read/write priorities suffice? Basically, we need to think about handling the case of "Cannot write to chosen shard as it's full" by migrating less-used blocks on that shard to lower-priority shards, and the case of "having to fetch blocks from low-read-priority shards as they don't exist elsewhere" by considering promoting them to higher read priority shards. This will require keeping some kind of time-averaged usage score for blocks in the metadata.

Other Links to this Post

RSS feed for comments on this post.

Leave a comment

WordPress Themes

Creative Commons Attribution-NonCommercial-ShareAlike 2.0 UK: England & Wales
Creative Commons Attribution-NonCommercial-ShareAlike 2.0 UK: England & Wales