Identity/AttachedServices/StorageServiceArchitecture

Summary

This is a working proposal for the backend storage architecture of PiCL server. It's based on a massively-sharded and cross-DC-replicated MySQL installation, and is far from final. All feedback welcome!

The goals are:

• Scale to billions of users. Quickly. Easily.
• Don't lose user data. Even if a machine dies. Even if a meteor hits a data-center.
• Maximize uptime, running costs be damned.
• Provide a simple programming model to the application.
• Provide a relatively simple and well-understood Ops environment.

The basic principles:

• Each user's data is completely independent, there's no need for queries that cross multiple user accounts.
  • This means that our data storage problem is embarrassingly shardable. Good times!

• Each user account is assigned to a particular shard, identified by an integer.
  • Their shard assignment will never change unless they delete and re-create their account.

• All reads and writes for a shard go to a single master MySQL database.
  • This saves us having to deal with conflicting writes and other multi-master headaches.
  • To keep things simple, there are no read slaves. The sharding is the only thing responsible for distributing server load.
  • A single DB host machine might hold multiple shards.

• Each master synchronously replicates to a hot standby in the same DC, to guard against individual machine failure.

• Each master asynchronously replicates to a warm standby in a separate DC, to guard against whole-DC failure.

• All sharding logic and management lives in a stand-alone proxy process, so that it's transparent to the application.

What the App Sees

From the POV of the application code, it's just talking to a regular old MySQL database:

   +---------+          +--------------+
   | Web App |--------->| MySQL Server |
   +---------+          +--------------+

It has to respect a couple of restrictions though:

• No DDL statements are allowed, only regular queries. Schema changes and sharding don't mix.
• All queries must specify a fixed userid, e.g. by having a "userid = XXX" component in the WHERE clause, or by inserting rows with a known userid. This enables us to do the sharding transparently.

Transparent Sharding Proxy

The application code is actually talking to a proxy server that speaks the MySQL wire protocol. In turn, the proxy is talking to the individual MySQL servers that are hosting each shard:

                                                  +---------------------------+
                                           +----->| MySQL Server for Shard #1 |
   +---------+        +----------------+   |      +---------------------------+
   | Web App |------->| Sharding Proxy |---+
   +---------+        +----------------+   |      +---------------------------+
                                           +----->| MySQL Server for Shard #2 |
                                                  +---------------------------+

The proxy process will:

• Receive, parse and validate each incoming query
• Extract the target userid, erroring out if the query does not have one.
• Look up the shard number and corresponding database for that userid.
• Forward the query to the appropriate database host, and proxy back the results.

The particulars of shard selection/lookup are not defined in this proposal, and are orthogonal to the rest of the setup. :rfkelly likes the consistent-hashing-plus-vbucket approach taken by couchbase, but it could be as simple as a lookup table. We assume that the proxy implements this appropriately and efficiently.

Handling all the sharding logic in a separate proxy process gives us a number of advantages:

• Application code is greatly simplified.
  • The same code paths are exercised in deployment, testing, and third-party deployments against a single database machine.

• The total number of connections is reduced, along with various connection-related overheads.

• The proxy can do centralized health monitoring of the individual servers.

Intra-DC Redundancy

We need to guard against the loss of any individual server within a DC. There are separate redundancy schemes for the MySQL servers, and for the other supporting services.

MySQL Redundancy

To guard against the loss of any individual database server, each shard will also have a hot standby database, living in the same DC and configured for synchronous (semi-synchronous?) replication. The proxy monitors the health of the standby database, but does not forward it any queries. Its only job is to serve as a backup for the active master:

                                                  +---------------------+
                                           +----->| Master for Shard #1 |
                                           |      +----------+----------+
                                           |                 | (replication)
   +---------+        +----------------+   |      +----------V---------------+
   | Web App |------->| Sharding Proxy |---+----->| Hot Standby for Shard #1 |
   +---------+        +----------------+   |      +--------------------------+
                                           |
                                           |      +---------------------+
                                           +----->| Master for Shard #2 |
                                           |      +----------+----------+
                                           |                 | (replication)
                                           |      +----------V---------------+
                                           +----->| Hot Standby for Shard #2 |
                                                  +--------------------------+

The proxy process is responsible for monitoring the health of these machines and sounding the alarm if something goes wrong. If the active master appears to be down, the proxy will transparently promote the hot standby and start sending queries to it. When the downed master comes back up, it is demoted to being the new standby.

TODO: The failover could be performed manually, if we're a bit leery of infrastructure being too clever for its own good.

TODO: Just one standby? Two? The principle should be the same regardless of how many we have. Star Topology FTW.

TODO: We could use the standby as a read slave, but I don't see the point. In a failure scenario the master needs to be able to handle the entire read load on its own, so it might as well do that all the time.

Other Service Redundancy

We don't want any single-point-of-failures, so we'll have to have multiple instances of the webapp talking to multiple instances of the proxy. These are connected via loadbalancing, virtual IPs, and whatever Ops wizardry is required to make single-machine failures in each tier be a non-event:

 +--------------+    +------------------+
 | Web App Tier |    | Shard Proxy Tier |         +---------------------+
 |              |    |                  |     +-->| Master for Shard #N |
 |  +---------+ |    | +-------------+  |     |   +----------+----------+
 |  | Web App | |--->| | Shard Proxy |  |-----+              | (replication)
 |  +---------+ |    | +-------------+  |     |   +----------V---------------+
 |  +---------+ |    | +-------------+  |     +-->| Hot Standby for Shard #N |
 |  | Web App | |    | | Shard Proxy |  |         +--------------------------+
 |  +---------+ |    | +-------------+  |
 +--------------+    +------------------+

Note that we're not doing this to the MySQL servers. There's too many of them and we already have a custom redundancy scheme from the hot standby.

With multiple Shard Proxy processes, we run into the problem of shared state. They must all agree on the current mapping of userids to shards, of shards to database machines, and which database machines are master versus standby. They'll operate as a ZooKeeper (or similar) cluster to store this state in a consistent and highly-available fashion:

  +----------------------------------------------+
  | Shard Proxy Tier                             |
  |                                              |
  |  +------------------+    +-----------------+ |
  |  | Shard Proxy:     |    | Shard Proxy:    | |
  |  |  ZooKeeper Node <+----+> ZooKeeper Node | |
  |  |  Proxy Process   |    |  Proxy Process  | |
  |  +----|-------------+    +----|------------+ |
  |       |                       |              |
  |       +-----------+-----------+              |
  +-------------------|--------------------------+
                      V
              ...................
              : MySQL Instances :
              :.................:

Note that this shard-state metadata will be very small and be updated very infrequently, which should make it very friendly to a local zookeeper installation.

Inter-DC Redundancy

We'll replicate the entire stack into several data-centers, each of which will maintain a full copy of all shards.

One DC will be the active master for each shard. All reads and writes for that shard will be forwarded into that DC and routed to the master. Other DCs are designated as warm-standby hosts for that shard, configured for asynchronous WAN replication. They can be failed-over to if there is a serious outage in the master DC, but this will almost certainly result in the loss of some recent transactions:

 +----------------------------------------------------------------------------------+
 | US-East Data Center                                                              |
 |                                                                                  |
 |   +--------------+    +------------------+                                       |
 |   | Web App Tier |    | Shard Proxy Tier |         +---------------------+       |
 |   |              |    |                  |     +-->| Master for Shard #N |-------+-----+
 |   |  +---------+ |    | +-------------+  |     |   +----------+----------+       |     |
 |   |  | Web App | |--->| | Shard Proxy |  |-----+              | (replication)    |     |
 |   |  +---------+ |    | +-------------+  |     |   +----------V---------------+  |     |
 |   |  +---------+ |    | +-------------+  |     +-->| Hot Standby for Shard #N |  |     | (very slow replication)
 |   |  | Web App | |    | | Shard Proxy |  |         +--------------------------+  |     |
 |   |  +---------+ |    | +-------------+  |                                       |     |
 |   +--------------+    +------------------+                                       |     |
 |                                |                                                 |     |
 +--------------------------------+-------------------------------------------------+     |
                                  |                                                       |
                                  | (very slow replication)                               |
                                  |                                                       |
 +--------------------------------+---------------------------------------------------+   |
 | US-West Data Center            |                                                   |   |
 |                                V                                                   |   |
 |   +--------------+    +------------------+                                         |   |
 |   | Web App Tier |    | Shard Proxy Tier |         +---------------------------+   |   |
 |   |              |    |                  |     +-->| Warm Standby for Shard #N |<--|---+
 |   |  +---------+ |    | +-------------+  |     |   +----------+----------------+   |
 |   |  | Web App | |--->| | Shard Proxy |  |-----+              | (replication)      |
 |   |  +---------+ |    | +-------------+  |     |   +----------V-----------------+  |
 |   |  +---------+ |    | +-------------+  |     +-->| Tepid Standby for Shard #N |  |
 |   |  | Web App | |    | | Shard Proxy |  |         +----------------------------+  |
 |   |  +---------+ |    | +-------------+  |                                         |
 |   +--------------+    +------------------+                                         |
 +------------------------------------------------------------------------------------+

For this scheme to work, every Shard Proxy process in every DC needs to agree on where the master is for every shard. That's not trivial. But it should be pretty doable with some replication between the respective ZooKeeper masters in each DC. Remember: this data is small and changes infrequently.

We sould probably not try to automate whole-DC failover, so that Ops can ensure consistent state before anything tries to send writes to a new location.

TODO: How many DCs? The principle should be the same regardless of how many we have. Nested Star Topology FTW.

TODO: Will we wind up in a situation where different shard have their master db in different DCs? Does it matter?

TODO: Do we want to try for any sort of geo-locality with data storage, e.g. putting the master for European users in Europe? This could greatly complicate the assignment of users to shards.

Implications

Using a single master means we don't have to worry about conflicts or consistency. The sharding means this should not be a bottle-neck, and the use of an intermediate proxy means we can fail over fast if the master goes down.

However, since we're doing asynchronous replication, there's a chance that recent database writes could be lost in the event of failure. The client will see a consistent, but out-of-date view of its data. It must be able to recover from such a situation, although we hope this would be a very rare occurrence!

Implementing the Proxy

The Sharding Proxy process is obviously key here, and looks like a reasonably complex beast. Can we use an off-the-shelf solution for this? There are some that have most of the required features, e.g. ScaleBase.

On the other hand, the feature set seems small enough that we could realistically implement the proxy in-house, with the benefit of tighter focus and greater control over the details of monitoring, failover, replication etc.

Things To Think About

I've tried to strike a balance between operational simplicity, application simplicity, and functionality here. We pay a price for it though:

• There's quite a few moving parts here. ZooKeeper is a beast. The proxy process has a few different, interacting responsibilities that would have to be carefully modeled and managed.

• There's the potential for severe cross-DC latency if you receive a HTTP request in one data-center, but have to forward all the MySQL queries over to the master in another data-center. I don't think there's any way around this without going to an eventually-consistent model, which would complicate the client API.

• There's a lot of redundancy here, which will cost a lot to run. Are our uptime requirements really so tight that we need a warm-standby in a separate DC? Could we get away with just the hot standby and periodic database dumps into S3, with which we can (slowly) recover from meteor-hit-the-data-center scale emergencies?

• How will we cope with moving shards between database hosts, or replacing dead hosts with fresh machines that have to catch up to the master. Checkpointing for faster recovery?