Thanks so much for your replies. This has been a great help understanding Rabbit better with having very little experience with it. I have a few follow up comments below.
> While you are correct the payload is a much bigger concern, managing the > metadata and acks centrally on the broker across multiple clients at scale > is also a concern. This would seem to be exasperated if you have consumers > at different speeds i.e. Storm and Hadoop consuming the same topic. > > In that scenario, say storm consumes the topic messages in real-time and > Hadoop consumes once a day. Let’s assume the topic consists of 100k+ > messages/sec throughput so that in a given day you might have 100s GBs of > data flowing through the topic. > > To allow Hadoop to consume once a day, Rabbit obviously can’t keep 100s GBs > in memory and will need to persist this data to its internal DB to be > retrieved later. I am not sure why you think this is a problem? For a fixed number of producers and consumers, the pubsub and delivery semantics of Rabbit and Kafka are quite similar. Think of Rabbit as adding an in-memory cache that is used to (a) speed up read consumption, (b) obviate disk writes when possible due to all client consumers being available and consuming. Actually I think this is the main use case that sets Kafka apart from Rabbit and speaks to the poster’s ‘Arguments for Kafka over RabbitMQ’ question. As you mentioned Rabbit is a general purpose messaging system and along with that has a lot of features not found in Kafka. There are plenty of times when Rabbit makes more sense than Kafka, but not when you are maintaining large message stores and require high throughput to disk. Persisting 100s GBs of messages to disk is a much different problem than managing messages in memory. Since Rabbit must maintain the state of the consumers I imagine it’s subjected to random data access patterns on disk as opposed to sequential. Quoting the Kafka design page ( http://kafka.apache.org/07/design.html) performance of sequential writes on a 6 7200rpm SATA RAID-5 array is about 300MB/sec but the performance of random writes is only about 50k/sec—a difference of nearly 10000X. They go on to say persistent data structure used in messaging systems metadata is often a BTree. BTrees are the most versatile data structure available, and make it possible to support a wide variety of transactional and non-transactional semantics in the messaging system. They do come with a fairly high cost, though: Btree operations are O(log N). Normally O(log N) is considered essentially equivalent to constant time, but this is not true for disk operations. Disk seeks come at 10 ms a pop, and each disk can do only one seek at a time so parallelism is limited. Hence even a handful of disk seeks leads to very high overhead. Since storage systems mix very fast cached operations with actual physical disk operations, the observed performance of tree structures is often superlinear. Furthermore BTrees require a very sophisticated page or row locking implementation to avoid locking the entire tree on each operation. The implementation must pay a fairly high price for row-locking or else effectively serialize all reads. Because of the heavy reliance on disk seeks it is not possible to effectively take advantage of the improvements in drive density, and one is forced to use small (< 100GB) high RPM SAS drives to maintain a sane ratio of data to seek capacity. Intuitively a persistent queue could be built on simple reads and appends to files as is commonly the case with logging solutions. Though this structure would not support the rich semantics of a BTree implementation, but it has the advantage that all operations are O(1) and reads do not block writes or each other. This has obvious performance advantages since the performance is completely decoupled from the data size--one server can now take full advantage of a number of cheap, low-rotational speed 1+TB SATA drives. Though they have poor seek performance, these drives often have comparable performance for large reads and writes at 1/3 the price and 3x the capacity. Having access to virtually unlimited disk space without penalty means that we can provide some features not usually found in a messaging system. For example, in kafka, instead of deleting a message immediately after consumption, we can retain messages for a relative long period (say a week). Our assumption is that the volume of messages is extremely high, indeed it is some multiple of the total number of page views for the site (since a page view is one of the activities we process). Furthermore we assume each message published is read at least once (and often multiple times), hence we optimize for consumption rather than production. There are two common causes of inefficiency: too many network requests, and excessive byte copying. To encourage efficiency, the APIs are built around a "message set" abstraction that naturally groups messages. This allows network requests to group messages together and amortize the overhead of the network roundtrip rather than sending a single message at a time. The MessageSet implementation is itself a very thin API that wraps a byte array or file. Hence there is no separate serialization or deserialization step required for message processing, message fields are lazily deserialized as needed (or not deserialized if not needed). The message log maintained by the broker is itself just a directory of message sets that have been written to disk. This abstraction allows a single byte format to be shared by both the broker and the consumer (and to some degree the producer, though producer messages are checksumed and validated before being added to the log). Maintaining this common format allows optimization of the most important operation: network transfer of persistent log chunks. Modern unix operating systems offer a highly optimized code path for transferring data out of pagecache to a socket; in Linux this is done with the sendfile system call. Java provides access to this system call with the FileChannel.transferTo api. To understand the impact of sendfile, it is important to understand the common data path for transfer of data from file to socket: 1. The operating system reads data from the disk into pagecache in kernel space 2. The application reads the data from kernel space into a user-space buffer 3. The application writes the data back into kernel space into a socket buffer 4. The operating system copies the data from the socket buffer to the NIC buffer where it is sent over the network This is clearly inefficient, there are four copies, two system calls. Using sendfile, this re-copying is avoided by allowing the OS to send the data from pagecache to the network directly. So in this optimized path, only the final copy to the NIC buffer is needed. We expect a common use case to be multiple consumers on a topic. Using the zero-copy optimization above, data is copied into pagecache exactly once and reused on each consumption instead of being stored in memory and copied out to kernel space every time it is read. This allows messages to be consumed at a rate that approaches the limit of the network connection. So in the end it would seem Kafka’s specialized nature to write data first really shines over Rabbit when your use case requires a very high throughput unblocking firehose with large data persistence to disk. Since this is only one use case this by no means is saying Kafka is better than Rabbit or vice versa. I think it is awesome there are more options to choose from so you can pick the right tool for the job. Thanks open source! As always YMMV. On Fri, Jun 7, 2013 at 4:40 PM, Alexis Richardson < alexis.richard...@gmail.com> wrote: > Jonathan, > > > On Fri, Jun 7, 2013 at 7:03 PM, Jonathan Hodges <hodg...@gmail.com> wrote: > > Hi Alexis, > > > > I appreciate your reply and clarifications to my misconception about > > Rabbit, particularly on the copying of the message payloads per consumer. > > Thank-you! > > > > It sounds like it only copies metadata like the consumer state i.e. > > position in the topic messages. > > Basically yes. Of course when a message is delivered to N>1 > *machines*, then there will be N copies, one per machine. > > Also, for various reasons, very tiny (<60b) messages do get copied as > you'd assumed. > > > > I don’t have experience with Rabbit and > > was basing this assumption based on Google searches like the following - > > > http://ilearnstack.com/2013/04/16/introduction-to-amqp-messaging-with-rabbitmq/ > . > > It seems to indicate with topic exchanges that the messages get copied > to > > a queue per consumer, but I am glad you confirmed it is just the > metadata. > > Yup. > > That's a fairly decent article but even the good stuff uses words like > "copy" without a fixed denotation. Don't believe the internets! > > > > While you are correct the payload is a much bigger concern, managing the > > metadata and acks centrally on the broker across multiple clients at > scale > > is also a concern. This would seem to be exasperated if you have > consumers > > at different speeds i.e. Storm and Hadoop consuming the same topic. > > > > In that scenario, say storm consumes the topic messages in real-time and > > Hadoop consumes once a day. Let’s assume the topic consists of 100k+ > > messages/sec throughput so that in a given day you might have 100s GBs of > > data flowing through the topic. > > > > To allow Hadoop to consume once a day, Rabbit obviously can’t keep 100s > GBs > > in memory and will need to persist this data to its internal DB to be > > retrieved later. > > I am not sure why you think this is a problem? > > For a fixed number of producers and consumers, the pubsub and delivery > semantics of Rabbit and Kafka are quite similar. Think of Rabbit as > adding an in-memory cache that is used to (a) speed up read > consumption, (b) obviate disk writes when possible due to all client > consumers being available and consuming. > > > > I believe when large amounts of data need to be persisted > > is the scenario described in the earlier posted Kafka paper ( > > > http://research.microsoft.com/en-us/um/people/srikanth/netdb11/netdb11papers/netdb11-final12.pdf > ) > > where Rabbit’s performance really starts to bog down as compared to > Kafka. > > Not sure what parts of the paper you mean? > > I read that paper when it came out. I found it strongest when > describing Kafka's design philosophy. I found the performance > statements made about Rabbit pretty hard to understand. This is not > meant to be a criticism of the authors! I have seen very few > performance papers about messaging that I would base decisions on. > > > > This Kafka paper is looks to be a few years old > > Um.... Lots can change in technology very quickly :-) > > Eg.: At the time this paper was published, Instagram had 5m users. > Six months earlier in Dec 2010, it had 1m. Since then it grew huge > and got acquired. > > > > > so has something changed > > within the Rabbit architecture to alleviate this issue when large amounts > > of data are persisted to the internal DB? > > Rabbit introduced a new internal flow control system which impacted > performance under steady load. This may be relevant? I couldn't say > from reading the paper. > > I don't have a good reference for this to hand, but here is a post > about external flow control that you may find amusing: > > http://www.rabbitmq.com/blog/2012/05/11/some-queuing-theory-throughput-latency-and-bandwidth/ > > > > Do the producer and consumer > > numbers look correct? If no, maybe you can share some Rabbit benchmarks > > under this scenario, because I believe it is the main area where Kafka > > appears to be the superior solution. > > This is from about one year ago: > > http://www.rabbitmq.com/blog/2012/04/25/rabbitmq-performance-measurements-part-2/ > > Obviously none of this uses batching, which is an easy trick for > increasing throughput. > > YMMV. > > Is this helping? > > alexis > > > > > Thanks for educating me on these matters. > > > > -Jonathan > > > > > > > > On Fri, Jun 7, 2013 at 6:54 AM, Alexis Richardson <ale...@rabbitmq.com > >wrote: > > > >> Hi > >> > >> Alexis from Rabbit here. I hope I am not intruding! > >> > >> It would be super helpful if people with questions, observations or > >> moans posted them to the rabbitmq list too :-) > >> > >> A few comments: > >> > >> * Along with ZeroMQ, I consider Kafka to be one of the interesting and > >> useful messaging projects out there. In a world of cruft, Kafka is > >> cool! > >> > >> * This is because both projects come at messaging from a specific > >> point of view that is *different* from Rabbit. OTOH, many other > >> projects exist that replicate Rabbit features for fun, or NIH, or due > >> to misunderstanding the semantics (yes, our docs could be better) > >> > >> * It is striking how few people describe those differences. In a > >> nutshell they are as follows: > >> > >> *** Kafka writes all incoming data to disk immediately, and then > >> figures out who sees what. So it is much more like a database than > >> Rabbit, in that new consumers can appear well after the disk write and > >> still subscribe to past messages. Instead, Rabbit which tries to > >> deliver to consumers and buffers otherwise. Persistence is optional > >> but robust and a feature of the buffer ("queue") not the upstream > >> machinery. Rabbit is able to cache-on-arrival via a plugin, but this > >> is a design overlay and not particularly optimal. > >> > >> *** Kafka is a client server system with end to end semantics. It > >> defines order to include processing order, and keeps state on the > >> client to do this. Group management is via a 3rd party service > >> (Zookeeper? I forget which). Rabbit is a server-only protocol based > >> system which maintains order on the server and through completely > >> language neutral protocol semantics. This makes Rabbit perhaps more > >> natural as a 'messaging service' eg for integration and other > >> inter-app data transfer. > >> > >> *** Rabbit is a general purpose messaging system with extras like > >> federation. It speaks many protocols, and has core features like HA, > >> transactions, management, etc. Everything can be switched on or off. > >> Getting all this to work while keeping the install light and fast, is > >> quite fiddly. Kafka by contrast comes from a specific set of use > >> cases, which are interesting certainly. I am not sure if Kafka wants > >> to be a general purpose messaging system, but it will become a bit > >> more like Rabbit if that is the goal. > >> > >> *** Both approaches have costs. In the case of Rabbit the cost is > >> that more metadata is stored on the broker. Kafka can get performance > >> gains by storing less such data. But we are talking about some N > >> thousands of MPS versus some M thousands. At those speeds the clients > >> are usually the bottleneck anyway. > >> > >> * Let me also clarify some things: > >> > >> *** Rabbit does NOT store multiple copies of the same message across > >> queues, unless they are very small (<60b, iirc). A message delivered > >> to >1 queue on 1 machine is stored once. Metadata about that message > >> may be stored more than once, but, at scale, the big cost is the > >> payload. > >> > >> *** Rabbit's vanilla install does store some index data in memory when > >> messages flow to disk. You can change this by using a plugin, but > >> this is a secret-menu undocumented feature. Very very few people need > >> any such thing. > >> > >> *** A Rabbit queue is lightweight. It's just an ordered consumption > >> buffer that can persist and ack. Don't assume things about Rabbit > >> queues based on what you know about IBM MQ, JMS, and so forth. Queues > >> in Rabbit and Kafka are not the same. > >> > >> *** Rabbit does not use mnesia for message storage. It has its own > >> DB, optimised for messaging. You can use other DBs but this is > >> Complicated. > >> > >> *** Rabbit does all kinds of batching and bulk processing, and can > >> batch end to end. If you see claims about batching, buffering, etc., > >> find out ALL the details before drawing conclusions. > >> > >> I hope this is helpful. > >> > >> Keen to get feedback / questions / corrections. > >> > >> alexis > >> > >> > >> > >> > >> > >> > >> > >> On Fri, Jun 7, 2013 at 2:09 AM, Marc Labbe <mrla...@gmail.com> wrote: > >> > We also went through the same decision making and our arguments for > Kafka > >> > where in the same lines as those Jonathan mentioned. The fact that we > >> have > >> > heterogeneous consumers is really a deciding factor. Our requirements > >> were > >> > to avoid loosing messages at all cost while having multiple consumers > >> > reading the same data at a different pace. On one side, we have a few > >> > consumers being fed with data coming in from most, if not all, > topics. On > >> > the other side, we have a good bunch of consumers reading only from a > >> > single topic. The big guys can take their time to read while the > smaller > >> > ones are mostly for near real-time events so they need to keep up the > >> pace > >> > of incoming messages. > >> > > >> > RabbitMQ stores data on disk only if you tell it to while Kafka > persists > >> by > >> > design. From the beginning, we decided we would try to use the queues > the > >> > same way, pub/sub with a routing key (an exchange in RabbitMQ) or > topic, > >> > persisted to disk and replicated. > >> > > >> > One of our scenario was to see how the system would cope with the > largest > >> > consumer down for a while, therefore forcing the brokers to keep the > data > >> > for a long period. In the case of RabbitMQ, this consumer has it owns > >> queue > >> > and data grows on disk, which is not really a problem if you plan > >> > consequently. But, since it has to keep track of all messages read, > the > >> > Mnesia database used by RabbitMQ as the messages index also grows > pretty > >> > big. At that point, the amount of RAM necessary becomes very large to > >> keep > >> > the level of performance we need. In our tests, we found that this an > >> > adverse effect on ALL the brokers, thus affecting all consumers. You > can > >> > always say that you'll monitor the consumers to make sure it won't > >> happen. > >> > That's a good thing if you can. I wasn't ready to make that bet. > >> > > >> > Another point is the fact that, since we wanted to use pub/sub with a > >> > exchange in RabbitMQ, we would have ended up with a lot data > duplication > >> > because if a message is read by multiple consumers, it will get > >> duplicated > >> > in the queue of each of those consumer. Kafka wins on that side too > since > >> > every consumer reads from the same source. > >> > > >> > The downsides of Kafka were the language issues (we are using mostly > >> Python > >> > and C#). 0.8 is very new and few drivers are available at this point. > >> Also, > >> > we will have to try getting as close as possible to once-and-only-once > >> > guarantee. There are two things where RabbitMQ would have given us > less > >> > work out of the box as opposed to Kafka. RabbitMQ also provides a > bunch > >> of > >> > tools that makes it rather attractive too. > >> > > >> > In the end, looking at throughput is a pretty nifty thing but being > sure > >> > that I'll be able to manage the beast as it grows will allow me to > get to > >> > sleep way more easily. > >> > > >> > > >> > On Thu, Jun 6, 2013 at 3:28 PM, Jonathan Hodges <hodg...@gmail.com> > >> wrote: > >> > > >> >> We just went through a similar exercise with RabbitMQ at our company > >> with > >> >> streaming activity data from our various web properties. Our use > case > >> >> requires consumption of this stream by many heterogeneous consumers > >> >> including batch (Hadoop) and real-time (Storm). We pointed out that > >> Kafka > >> >> acts as a configurable rolling window of time on the activity stream. > >> The > >> >> window default is 7 days which allows for supporting clients of > >> different > >> >> latencies like Hadoop and Storm to read from the same stream. > >> >> > >> >> We pointed out that the Kafka brokers don't need to maintain consumer > >> state > >> >> in the stream and only have to maintain one copy of the stream to > >> support N > >> >> number of consumers. Rabbit brokers on the other hand have to > maintain > >> the > >> >> state of each consumer as well as create a copy of the stream for > each > >> >> consumer. In our scenario we have 10-20 consumers and with the scale > >> and > >> >> throughput of the activity stream we were able to show Rabbit quickly > >> >> becomes the bottleneck under load. > >> >> > >> >> > >> >> > >> >> On Thu, Jun 6, 2013 at 12:40 PM, Dragos Manolescu < > >> >> dragos.manole...@servicenow.com> wrote: > >> >> > >> >> > Hi -- > >> >> > > >> >> > I am preparing to make a case for using Kafka instead of Rabbit MQ > as > >> a > >> >> > broker-based messaging provider. The context is similar to that of > the > >> >> > Kafka papers and user stories: the producers publish monitoring > data > >> and > >> >> > logs, and a suite of subscribers consume this data (some store it, > >> others > >> >> > perform computations on the event stream). The requirements are > >> typical > >> >> of > >> >> > this context: low-latency, high-throughput, ability to deal with > >> bursts > >> >> and > >> >> > operate in/across multiple data centers, etc. > >> >> > > >> >> > I am familiar with the performance comparison between Kafka, > Rabbit MQ > >> >> and > >> >> > Active MQ from the NetDB 2011 paper< > >> >> > > >> >> > >> > http://research.microsoft.com/en-us/um/people/srikanth/netdb11/netdb11papers/netdb11-final12.pdf > >> >> >. > >> >> > However in the two years that passed since then the number of > >> production > >> >> > Kafka installations increased, and people are using it in different > >> ways > >> >> > than those imagined by Kafka's designers. In light of these > >> experiences > >> >> one > >> >> > can use more data points and color when contrasting to Rabbit MQ > >> (which > >> >> by > >> >> > the way also evolved since 2011). (And FWIW I know I am not the > first > >> one > >> >> > to walk this path; see for example last year's OSCON session on the > >> State > >> >> > of MQ<http://lanyrd.com/2012/oscon/swrcz/>.) > >> >> > > >> >> > I would appreciate it if you could share measurements, results, or > >> even > >> >> > anecdotal evidence along these lines. How have you avoided the > "let's > >> use > >> >> > Rabbit MQ because everybody else does it" route when solving > problems > >> for > >> >> > which Kafka is a better fit? > >> >> > > >> >> > Thanks, > >> >> > > >> >> > -Dragos > >> >> > > >> >> > >> >
