Someone's bound to say "But it doesn't support RPC!"
No, it doesn't, and that's a good thing. Use ZeroMQ as your transport with a protobuff payload and do remote message passing* instead of RPC - it'll scale more efficiently than straight RPC and you won't end up having to pool connections or queue requests in your client.
*Give each message request an identifier and have each response identify what message it's answering.
A research project with which I am involved recently moved from google protocol buffers + TCP to thrift RPC, because they found that thrift was much faster. Google Protocol buffers encode the keys and field types with varints, which are basically series of bytes of arbitrary length to represent ints. If the 8th bit is set, move on to the next byte. It's pretty straightforward
int parse(uchar *msg) {
int x = 0;
do {
x <<= 7;
x |= *msg;
} while(*msg & 0x80);
return x;
}
Thrift, on the other hand, uses vanilla 4-byte ints for the same purpose. This has the effect of limited the number of fields to 2^29 (I think they use the bottom 3 bits to specify field type, but I don't remember off the top of my head). On the flip side, it's pretty instant to parse, as opposed to the looping thing.
Further, thrift makes it pretty trivial to write a thrift server/client in any language.
In short, thrift > protocol buffers.
Edit: One caveat -- Async calls WITH return values are more or less a no-go in thrift.
I suspect that this is heavily dependent upon the relative speeds of network vs. CPU, as well as the usage that these encoded messages will get afterwards. Protocol buffers are making a speed vs. space tradeoff, minimizing the size of a message at the cost of some processing power. That probably made a lot more sense when Google was young and 100BaseT was new than in the era with ubiquitous gigabit Ethernet that we've got now.
Also remember that for most production services, you'll want to log requests & responses. As soon as you start hitting the disk, increased space usage gets expensive pretty fast. You'd probably want to convert your 32-bit ints to varints then anyway, and then you've lost all the CPU that you saved in the first place.
In short, YMMV. Consider your particular use-case carefully before making any blanket judgments of thrift > protobufs or protobufs > thrift.
You're talking about saving 3 bytes per int. I'd imagine it's pretty hard to find a workload where you can achieve 75% compression by using exclusively varints that are less than 128 in magnitude. In fact, I'd imagine that most of the data shipped in most workloads is mostly strings/binary, which have very little if any savings.
So let's talk about the ideal situation where we have exclusively fields that are varints that are limited to 128. This means that each will be 2 bytes (1 for field data + 1 for the varint itself, both of which are under 128 in magnitude). How many of these can we fit? Well, if we use the bottom 3 bits for specifying the type (a conservative estimate) that leaves only 4 bits (remember, subtract 1 for varint or not) to ID them, which means 16. So if a message has 16 fields and we save 6 bytes per request, we're talking a savings of 96 bytes. For the next 2048 entries, we're saving, 5 bytes per (5kb vs 15kb), then beyond that, we're saving 4 bytes per (and eventually none). This is also an incredibly arbitrary situation where you expect ALL of your fields to be small ints.
Same with disk, it's unlikely that your workloads will allow you to save a lot by using varints.
So, you're right that, for some workloads, this may be better, but if someone put a gun to my head and said "protocol buffers or thrift?" I'd have to go with thrift.
Going to reply to all your comments at this rate. ;-)
Even with single-byte tags, the overhead can be quite significant; thus the packed = true option. Unfortunately I don't have any public protocols or data to show you, but by switching to packed = true (combined with some extra tricks) I achieved ~40-60% savings on some of our datasets. We use protobufs for our stable storage, and this change certainly saved us more than my yearly compensation package.
> For the next 2048 entries
Yikes! I hope you never have to work with a message type with anywhere near 2048 fields! :-)
In practice, more than half message types have less than 16 fields, which is why the "incredibly arbitrary situation" is justified, and really does end up saving a ton of storage. But, for simplicity, nothing can beat just ntohl and a cast.
Have you run benchmarks on this? If so, can you link me to them?
It would be relatively easy to modify protobuf's parsing code to operate on the encoded byte rather than decoding field tags. If it's worth doing I'll do it.
BTW, the default parsing code has no loops in the fast-path. I don't remember off the top of my head, but I think both 1 and 2 byte cases are inlined. (Avoiding the loops is great for these cases, since default branch prediction will often be wrong.) All tags under 16 are one byte.
Also, in my experience IO has always been the bottleneck -- my CPU can easily keep up with 1 or 10 gig pipes I have access to.
You raise valid points, so I'll elaborate on my thinking: I don't really care what binary serialization format is used (personally, I use thrift), there's no shortage of good ones available (msgpack, thrift, protobuffs, avro, etc). I think it's more important to decouple the serialization format from the transport and protocol, so that when/if you realize you need to switch you only have to replace one piece.
I use rabbitmq for my android app.
It's just a matter of sending a byte array between the client and the server.
Then it's just a matter of reading the bytes into the appropriate fields. (both server and client)
This may be a bit of work but it's really fast and efficient.
After several releases of your app, you'll have several wire formats in the wild. One of the main features of libraries like Protocol Buffers is that they can parse old messages, so you don't need to maintain an increasingly hairier parsing routine.
Well, I deny access to the old clients. I consider this an advantage because I force users to keep up to date.
If you're careful you can just add items to the end of the stream and the old clients will work just fine. I don't
support this though, but I do it in practice.
But yeah, protobuf wins in that domain. Specially if it avoids a high serialization/deserialization overhead penalty. It'd be cool to use rabbitmq to route the protobuf objects around with a cleaner interface.
Protocol buffers are great, but the Python implementation provided by Google is written in native Python and very slow for serializing/deserializing protobufs (see http://news.ycombinator.com/item?id=767882).
I've heard you can get around this by SWIG-wrapping the C++ implementation, but I haven't gotten around to trying it yet.
I've recently moved to protocol buffers and am quite happy with the results. Before that, we had a few tomcat apps with ad-hoc urls that return xml responses. Since we already have a good http request/response channel, I was able to create a java bridge for the message objects, as well as a simple python library that can communicate with all our servers.
While there was a little overhead in getting it integrated into the build system, it works great, and gives a much better interface, and much better control for messaging individual servers.
The project itself is also very well put together. Hell, even the autoconf generated makefile has targets like "uninstall", which seems to be a rarity when building packages from source.
Both protocol buffers and thrift are primarily designed to handle older or newer messages (than your program was compiled with) gracefully. Speed and compactness appear to be secondary goals. (Especially for Thrift, which does not seem concerned with compactness.)
Rather than version your protocol, the generated code is able to parse newer messages. If you modify a field on that message and serialize it, even the parts of the message your binary is unaware of will be retained. (As opposed to dropping them.) This means you don't need to push all changes at once, and you can have binaries talk with older and new binaries without problems.
The basic wire format is [(tag, payload bytes)*]. The tag identifies both the field "number", and the wire-type of the field. Wire types include fixed 32 and 64, length-delimited ranges, and, for protos, variable length integers. The parser knows how to skip fields of all wire types, so any unknown tags are skipped (and retained in memory) during parsing.
Some caveats apply -- you can modify your protocol in several backwards-incompatible ways. For example: add a new required field. Your new binary will reject all serialized messages from your older binaries. Alternatively, change the type of a field. These are pretty easy to avoid, but given enough people on your team, it is bound to happen once or twice.
A great use case for Protocol Buffers or Thrift is adding schema to other technologies which just use raw byte arrays, like Hadoop files, Cassandra columns, and Gearman parameters
24 comments
[ 4.4 ms ] story [ 69.7 ms ] threadNo, it doesn't, and that's a good thing. Use ZeroMQ as your transport with a protobuff payload and do remote message passing* instead of RPC - it'll scale more efficiently than straight RPC and you won't end up having to pool connections or queue requests in your client.
*Give each message request an identifier and have each response identify what message it's answering.
Further, thrift makes it pretty trivial to write a thrift server/client in any language.
In short, thrift > protocol buffers.
Edit: One caveat -- Async calls WITH return values are more or less a no-go in thrift.
Also remember that for most production services, you'll want to log requests & responses. As soon as you start hitting the disk, increased space usage gets expensive pretty fast. You'd probably want to convert your 32-bit ints to varints then anyway, and then you've lost all the CPU that you saved in the first place.
In short, YMMV. Consider your particular use-case carefully before making any blanket judgments of thrift > protobufs or protobufs > thrift.
So let's talk about the ideal situation where we have exclusively fields that are varints that are limited to 128. This means that each will be 2 bytes (1 for field data + 1 for the varint itself, both of which are under 128 in magnitude). How many of these can we fit? Well, if we use the bottom 3 bits for specifying the type (a conservative estimate) that leaves only 4 bits (remember, subtract 1 for varint or not) to ID them, which means 16. So if a message has 16 fields and we save 6 bytes per request, we're talking a savings of 96 bytes. For the next 2048 entries, we're saving, 5 bytes per (5kb vs 15kb), then beyond that, we're saving 4 bytes per (and eventually none). This is also an incredibly arbitrary situation where you expect ALL of your fields to be small ints.
Same with disk, it's unlikely that your workloads will allow you to save a lot by using varints.
So, you're right that, for some workloads, this may be better, but if someone put a gun to my head and said "protocol buffers or thrift?" I'd have to go with thrift.
Even with single-byte tags, the overhead can be quite significant; thus the packed = true option. Unfortunately I don't have any public protocols or data to show you, but by switching to packed = true (combined with some extra tricks) I achieved ~40-60% savings on some of our datasets. We use protobufs for our stable storage, and this change certainly saved us more than my yearly compensation package.
> For the next 2048 entries
Yikes! I hope you never have to work with a message type with anywhere near 2048 fields! :-)
In practice, more than half message types have less than 16 fields, which is why the "incredibly arbitrary situation" is justified, and really does end up saving a ton of storage. But, for simplicity, nothing can beat just ntohl and a cast.
It would be relatively easy to modify protobuf's parsing code to operate on the encoded byte rather than decoding field tags. If it's worth doing I'll do it.
BTW, the default parsing code has no loops in the fast-path. I don't remember off the top of my head, but I think both 1 and 2 byte cases are inlined. (Avoiding the loops is great for these cases, since default branch prediction will often be wrong.) All tags under 16 are one byte.
Also, in my experience IO has always been the bottleneck -- my CPU can easily keep up with 1 or 10 gig pipes I have access to.
I think Thrift supports a lot more languages, but based on benchmarks, I believe it does perform (sometimes significantly) slower
I've never used Hessian though, will have to check it out at some point.
It has it's own rpc mechanism as well that competes with thrift, although it is not IDL-based (however, their IDL is coming soon).
But yeah, protobuf wins in that domain. Specially if it avoids a high serialization/deserialization overhead penalty. It'd be cool to use rabbitmq to route the protobuf objects around with a cleaner interface.
Shameless plug for my dev blog: http://developingthedream.blogspot.com/
I've heard you can get around this by SWIG-wrapping the C++ implementation, but I haven't gotten around to trying it yet.
While there was a little overhead in getting it integrated into the build system, it works great, and gives a much better interface, and much better control for messaging individual servers.
The project itself is also very well put together. Hell, even the autoconf generated makefile has targets like "uninstall", which seems to be a rarity when building packages from source.
Rather than version your protocol, the generated code is able to parse newer messages. If you modify a field on that message and serialize it, even the parts of the message your binary is unaware of will be retained. (As opposed to dropping them.) This means you don't need to push all changes at once, and you can have binaries talk with older and new binaries without problems.
The basic wire format is [(tag, payload bytes)*]. The tag identifies both the field "number", and the wire-type of the field. Wire types include fixed 32 and 64, length-delimited ranges, and, for protos, variable length integers. The parser knows how to skip fields of all wire types, so any unknown tags are skipped (and retained in memory) during parsing.
Some caveats apply -- you can modify your protocol in several backwards-incompatible ways. For example: add a new required field. Your new binary will reject all serialized messages from your older binaries. Alternatively, change the type of a field. These are pretty easy to avoid, but given enough people on your team, it is bound to happen once or twice.