Photosrv (pronounced "photo-serve") - sharded reverse proxy HTTP server
Photosrv is a consistently hashed HTTP router aimed at distributing static content.
It allows you to configure dozens of photosrv daemons and hundreds of underlying nginx storage nodes, capable of storing and serving billions of files with low- latency (100ms or under with sufficient hardware). The system has been employed in conjunction with multiple CDN(s).
You can effectively consider it a load balancer for your static content.
Clients write (PUT/DELETE) to the storage nodes. Photosrv reverse proxies the incoming (GET/HEAD) requests to the storage node(s).
You define a small Apache-like config file, with a list of data centers (clusters). You must have at least one data center, but it could consist of machines all over the world, for instance.
Each data center contains a list of shards, numbered in order. Each shard is a list of nodes. If you want 3 copies of every file you can configure every shard to have 3 nodes. Each node should be a single individual machine.
The Photosrv router uses the config file in combination with a roster file to determine where to route. The roster file is a (constantly updated) list of active storage hosts.
For every incoming request, the router consistently hashes the incoming path to figure out which shard contains that file. The hashing is configurable; Jump hash is presently recommended.
Photosrv has a couple of configurable routing modes:
-
Forward to a random node within the calculated shard
-
A 'multi-HEAD timeout' mode which forwards to a successful HEAD node.
NOTE: The multi-HEAD timeout results in high read amplification, and requires solid networking infrastructure. The idea is that all nodes are asked 'HEAD /PATH', and a random "fastest" node within the timeout window will be forwarded the request. If every node responds with a 404 or a timeout, then we forward the request to a random node anyway.
Each storage node is a single logical machine running a WebDAV server that a Photosrv node can route requests to for reading files. The only current requirements are HEAD, GET, PUT, and DELETE. Help with implementing PATCH forwarding would be appreciated, to allow for resumable file upload support via tus. See CONTRIBUTING.md
Because Photosrv itself relies on the built-in Go http and reverse proxy packages, it has implicit support for streaming data and resumable downloads. However, it is up to the client to support resumable downloads as well.
Storing files is currently possible using Go and a sample command-line utility for uploading data via Go is provided in 'cmd/shardsprites'. The Perl client needs polish before it will be suitable for general-purpose consumption.
Photosrv may eventually handle PUTs and do replication itself, but currently it is up to the client to do that.
You have to manage your own metadata, i.e. track your own files.
Replication is currently implemented in the clients - not the server/cluster.
There is tooling provided for manual node repair. A gossip-based automated node repair and metadata storage mechanism could perhaps be developed.
You can not 'ls somebucket' and get a list of existing files. This works more like a distributed key-value store for arbitrarily sized blobs, where you just rely on HTTP and the filesystem/hardware underneath. You only have set(), get(), and delete(). (Well, HEAD).
Automatic rebalancing and migration are not yet supported. Manual modes have been executed but are not yet documented.
While you could, in theory, do everything on 1 node; A minimal cluster would be:
- At least 2 WebDAV storage hosts
- At least 1 Photosrv host
For a production setup, nginx is recommended. For a Gopher or hobbyist setup, you may prefer pocketdav. (See github.com/rbastic/pocketdav)
See ROADMAP.
See CONTRIBUTING.md
Look at ROADMAP, look at TODO, pick something and see if it exists in the 'Issues' section on github already, and announce that you're intending on working on it if so or maybe the person who was working on it needs help too!
This software was originally developed at Booking.com. With approval from Booking.com, this software was released as open source, for which the authors would like to express their gratitude.
See LICENSE file.