Im relatively new to activemq and one of the first things im trying to do is publish from a server process to 5,000 topics ( one topic per stock ). The server and broker manage to keep up fine.
However on the consumer side its very odd. If I subscribe to all 5k topics with a single wildcard consumer ( "mytopic.>" ) everything keeps up fine. However if I try to subscribe with a single consumer per topic the performance just drops out and it cant keep up.
Ive tried playing with prefetch limits and optimized ack modes, nothing seems to help.
Any idea why a single wildcard would be able to perform fine where as 5k individual topics would not?
I could just as well demultiplex the msgs myself but would expect activemq to be able to do this for me as efficiently as I can.
EDIT: Some more information and updates on this:
I was testing this on ~ 6,000 topics publishing once per second
I'm using the activemq-cpp c++ library, and I was creating 1 session for all topics. It turns out the activemq implementation is horribly inefficient, it does a linear scan of all topics on every messages ( twice actually ) when delivering a message to a session.
To make matters worse, if you create a session per topic it tries to create a thread per session so that blows out pretty quick.
But wait! There's an option on the connection, setAlwaysSessionAsync, so sessions dont create their own threads, great!
D'OH! not so fast, sessions still create some some RW mutex in non-async mode, and they use some home-grown TLS data which had a hardcoded limit of ~300 instances per thread... ugh
Ok, so I had to limit number of sessions i can create to ~ 150 ( I guess other objects are using TLS data as well ) and then round-robin my topics on these...
It would be nice if I can control how many threads I can have processing the data off the wire, but alas thats not exposed either.... ugh, hardcoded in activemq-cpp code
TLDR; activemq is a streaming pile of messy poo
Related
I'm looking to solve a problem that I have with the FIFO nature of messaging severs and queues. In some cases, I'd like to distribute the messages in a queue to the pool of consumers on a criteria other than the message order it was delivered in. Ideally, this would prevent users from hogging shared resources in the system. Take this overly simplified scenario:
There is a feature within an application where a user can empty their trash can.
This event dispatches a DELETE message for each item in trash can
The consumers for this queue invoke a web service that has a rate limited API.
Given that each user can have very large volumes of messages in their trash can, what options do we have to allow concurrent processing of each trash can without regard to the enqueue time? It seems to me that there are a few obvious solutions:
Create a separate queue and pool of consumers for each user
Randomize the message delivery from a single queue to a single pool of consumers
In our case, creating a separate queue and managing the consumers for each user really isn't practical. It can be done but I think I really prefer the second option if it's reasonable. We're using RabbitMQ but not necessarily tied to it if there is a technology more suited to this task.
I'm entertaining the idea of using Rabbit's message priorities to help randomize delivery. By randomly assigning a message a priority between 1 and 10, this should help distribute the messages. The problem with this method is that the messages with the lowest priority may be stuck in the queue forever if the queue is never completely emptied. I thought I could use a TTL on the message and then re-queue the message with an escalated priority but I noticed this in the docs:
Messages which should expire will still only expire from the head of
the queue. This means that unlike with normal queues, even per-queue
TTL can lead to expired lower-priority messages getting stuck behind
non-expired higher priority ones. These messages will never be
delivered, but they will appear in queue statistics.
I fear that I may heading down the rabbit hole with this approach. I wonder how others are solving this problem. Any feedback on creative routing, messaging patterns, or any alternative solutions would be appreaciated.
So I ended up taking a page out of the network router handbook. This a problem they routers need to solve to allow fair traffic patterns. This video has a good breakdown of the problem and the solution.
The translation of the problem into my domain:
And the solution:
The load balancer is a wrapper around a channel and a known number of queues that uses a weighted algorithm to balance between messages received on each queue. We found a really interesting article/implementation that seems to be working well so far.
With this solution, I can also prioritize workspaces after messages have been published to increase their throughput. That's a really nice feature.
The biggest challenge ahead of me is management of the queues. There will be too many queues to leave bound to the exchange for an extended period of time. I'm working on some tools to manage their lifecycle.
One solution could be to interpose a Resequencer. The principle is outlined in the diag in that link. In your case, something like:
The app dispatches its DELETE messages into the delete queue as originally.
The Resequencer (a new component you write) is interposed between the original publishers and original consumers. It:
pulls messages off the DELETE queue into memory
places them into (in-memory) queues-by-user
republishes them to a new queue (eg FairPriorityDeleteQueue), round-robinning to interleave fairly any messages from different original users
limits its republish rate into FairPriorityDeleteQueue, either such that the length of FairPriorityDeleteQueue (obtainable via polling the rabbitmq management api periodically) never exceeds some integer you choose N, or limited to some rate related to the rate-limited delete API the consumers use.
doesn't ack any message it pulled off the original DELETE queue, until it's republished it to FairPriorityDeleteQueue (so you never lose a message)
The original consumers subscribe instead to FairPriorityDeleteQueue.
You set the preFetchCount on these consumers fairly low (<10), to prevent them in turn bulk-buffering the contents of FairPriorityDeleteQueue in memory.
--
Some points to watch:
Rate- or length-limiting publishing into and/or drawing messages out of FairPriorityDeleteQueue is essential. If you don't limit, Resequencer may just hand messages on as fast as it receives them, limiting the potential for resequencing.
Resequencer of course acts as a kind of in-memory buffer while resequencing. If the original publishers can publish very large numbers of messages in to the queue suddenly, you may need to memory-limit the Resequencer process so that it doesn't ingest more than it can hold.
Your particular scenario is greatly helped by the fact that you have an external factor (the final delete API) limiting throughput. Without such an extrinsic limiting factor, it is much harder to choose the optimum parameters for such a resequencer, to balance throughput-versus-resequencing in a particular environment.
I don't think a resequencer is needed in this case. Maybe it is, if you need to ensure the items are deleted in a specific order. But that only comes into play when you send multiple messages at roughly the same time and need to guarantee order on the consumer end.
You should also avoid the timeout scenario, for the reasons you've mentioned. timeout is meant to tell RabbitMQ that a message doesn't need to be processed - or that it needs to be routed to a dead letter queue so that i can be processed by some other code. while you might be able to make timeout work, i don't think it's a good choice.
Priorities may solve part of the problem, but could introduce a scenario where files never get processed. if you have a priority 1 message sitting back in the queue somewhere, and you keep putting priority 2, 3, 5, 10, etc. into the queue, the 1 might not be processed. the timeout doesn't solve this, as you've noted.
For my money, I would suggest a different approach: sending delete requests serially, for a single file.
that is, send 1 message to delete 1 file. wait for a response to say it's done. then send the next message to delete the next file.
here's why i think that will work, and how to manage it:
Long-Running Workflow, Single File Delete Requests
In this scenario, I would suggest taking a multi-step approach to the problem using the idea of a "saga" (aka a long-running workflow object).
when a user requests to delete their trashcan, you send a single message through rabbitmq to the service that can handle the delete process. that service creates an instance of the saga for that user's trashcan.
the saga gathers a list of all files in the trashcan that need to be deleted. then it starts to send the requests to delete the individual files, one at a time.
with each request to delete a single file, the saga waits for the response to say the file was deleted.
when the saga receives the message to say the previous file has been deleted, it sends out the next request to delete the next file.
once all the files are deleted, the saga updates itself and any other part of the system to say the trash can is empty.
Handling Multiple Users
When you have a single user requesting a delete, things will happen fairly quickly for them. they will get their trash emptied soon.
u1 = User 1 Trashcan Delete Request
|u1|u1|u1|u1|u1|u1|u1|u1|u1|u1done|
when you have multiple users requesting a delete, the process of sending one file delete request at a time means each user will have an equal chance of getting the next file delete.
u1 = User 1 Trashcan Delete Request
u2 = User 2 Trashcan Delete Request
|u1|u2|u1|u1|u2|u2|u1|u2|u1|u2|u2|u1|u1|u1|u2|u2|u1|u2|u1|u1done|u2|u2done|
This way, there will be shared use of the resources to delete the files. Over-all, it will take a little longer for each person's trashcan to be emptied, but they will see progress sooner and that's an important aspect of people thinking the system is fast / responsive to their request.
Optimizing Small File Set vs Large File Set
In a scenario where you have a small number of users with a small number of files, the above solution may prove to be slower than if you deleted all the files at once. after all, there will be more messages sent across rabbitmq - at least 2 for every file that needs to be deleted (one delete request, one delete confirmation response)
To optimize this further, you could do a couple of things:
have a minimum trashcan size before you split up the work like this. below that minimum, you just delete it all at once
chunk the work into groups of files, instead of one at a time. maybe 10 or 100 files would be a better group size, than 1 file at a time
Either (or both) of these solutions would help to improve the over-all performance of the process by reducing the number of messages being sent, and batching the work a bit.
You would need to do some testing in your real scenario to see which of these (or maybe both) would help and at what settings.
Many Users Problem
There's one additional problem you may face - many users. If you have 2 or 3 users requesting deletes, it won't be a big deal.
But if you have 100 or 1000 users requesting deletes, it could take a very long time for an individual to get their trashcan emptied.
You may need to have a higher level controlling process for this situation, where all requests to empty trashcans would be managed by yet another Saga. This saga would rate-limit the number of active trashcan-deletion sagas.
For example, if you have 10 active requests for deleting trashcans, the rate-limiting saga would only start 3 of them and it would wait for one to finish before starting the next one.
Again, you would need to test your actual scenario to see if this is needed and see what the limits should be, for performance reasons.
There may be additional scenarios that have to be considered in your actual scenario, but I hope this gets you down the path! :)
This is a new area for me so hopefully my question makes sense.
In my program I have a large number of clients which are windows services running on laptops - that are often disconnected. Occasionally they come on line and I want them to receive updates based on user profiles. There are many types of notifications that require the client to perform some work on the local application (i.e. the laptop).
I realize that I could do this with a series of restful database queries, but since there are so many clients (upwards to 10,000) and there are lots of different notification types, I was curious if perhaps this was not a problem better suited for a messaging product like RabbitMQ or even 0MQ.
But how would one set this up. (let's assume in RabbitMQ?
Would each user be assigned their own queue?
Or is it preferable to have each queue be a distinct notification type and you would use some combination of direct exchanges or filtering messages based on a routing key, where the routing key could be a username.
Since each user may potentially have a different set of notifications based on their user profile, I am thinking that each client/consumer would have a specific message for each notification sitting on a queue waiting for them to come online and process it.
Is this the right way of thinking about the problem? Thanks in advance.
It will be easier for you to balance a lot of queues than filter long ones, so it's better to use queue per consumer.
Messages can have arbitrary headers and body so it is the right place for notification types.
Since you will be using long-living queues, waiting for consumers on disk - you better use lazy queues https://www.rabbitmq.com/lazy-queues.html (it's available since version 3.6.0)
We're currently using RabbitMQ, where a continuously super-fast producer is paired with a consumer limited by a limited resource (e.g. slow-ish MySQL inserts).
We don't like declaring a queue with x-max-length, since all messages will be dropped or dead-lettered once the limit is reached, and we don't want to loose messages.
Adding more consumers is easy, but they'll all be limited by the one shared resource, so that won't work. The problem still remains: How to slow down the producer?
Sure, we could put a flow control flag in Redis, memcached, MySQL or something else that the producer reads as pointed out in an answer to a similar question, or perhaps better, the producer could periodically test for queue length and throttle itself, but these seem like hacks to me.
I'm mostly questioning whether I have a fundamental misunderstanding. I had expected this to be a common scenario, and so I'm wondering:
What is best practice for throttling producers? How is this done with RabbitMQ? Or do you do this in a completely different way?
Background
Assume the producer actually knows how to slow himself down with the right input. E.g. a hardware sensor or hardware random number generator, that can generate as many events as needed.
In our particular real case, we have an API that users can use to add messages. Instead of devouring and discarding messages, we'd like to apply back-pressure by having our API return an error if the queue is "full", so the caller/user knows to back-off, or have the API block until the consumer catches up. We don't control our user, so regardless of how fast the consumer is, I can create a producer that is faster.
I was hoping for something like the API for a TCP socket, where a write() can block and where a select() can be used to determine if a handle is writable. So either having the RabbitMQ API block or have it return an error if the queue is full.
For the x-max-length property, you said you don't want messages to be dropped or dead-lettered. I see there was an update in adding some more capabilities for this. As I see it is specified in the documentation:
"Use the overflow setting to configure queue overflow behaviour. If overflow is set to reject-publish, the most recently published messages will be discarded. In addition, if publisher confirms are enabled, the publisher will be informed of the reject via a basic.nack message"
So as I understand it, you can use queue limit to reject the new messages from publishers thus pushing some backpressure to the upstream.
I don't think that this is in any way rabbitmq specific. Basically you have a scenario, where there are two systems of different processing capabilities, and this mismatch will either pose a risk of overflowing the queue (whatever it would be), or even in case of a constant mismatch between producer and consumer, simply create more and more time-distance between event creation and its handling.
I used to deal with this kind of scenarios, and unfortunately there is no magic bullet. You either have to speed up even handling (better hardware, more suited software?) or throttle the event creation (which has nothing to do with MQ really).
Now, I would ask you what's the goal and how the events are produced. Are the events are produced constantly, with either unlimitted or just very high rate (for example readings from sensors - the more, the better), or are they created in batches/spikes (for example: user requests in specific time periods, batch loads from CRM system). I assume that the goal is to process everything cause you mention you don't want to loose any queued message.
If the output is constant, then some limiter (either internal counter, if the producer is the only producer, or external queue length checks if queue can be filled with some other system) is definitely in place.
IF eventsInTimePeriod/timePeriod > estimatedConsumerBandwidth
THEN LowerRate()
ELSE RiseRate()
In real world scenarios we used to simply limit the output manually to the estimated values and there were some alerts set for queue length, time from queue entry to queue leaving etc. Where such limiters were omitted (by mistake mostly) we used to find later some tasks that were supposed to be handled in few hours, that were waiting for three months for their turn.
I'm afraid it's hard to answer to "How to slow down the producer?" if we know nothing about it, but some ideas are: aforementioned rate check or maybe a blocking AddMessage method:
AddMessage(message)
WHILE(getQueueLength() > maxAllowedQueueLength)
spin(1000); // or sleep or whatever
mqAdapter.AddMessage(message)
I'd say it all depends on specific of the producer application and in general your architecture.
I have a middleware based on Apache Camel which does a transaction like this:
from("amq:job-input")
to("inOut:businessInvoker-one") // Into business processor
to("inOut:businessInvoker-two")
to("amq:job-out");
Currently it works perfectly. But I can't scale it up, let say from 100 TPS to 500 TPS. I already
Raised the concurrent consumers settings and used empty businessProcessor
Configured JAVA_XMX and PERMGEN
to speed up the transaction.
According to Active MQ web Console, there are so many messages waiting for being processed on scenario 500TPS. I guess, one of the solution is scale the ActiveMQ up. So I want to use multiple brokers in cluster.
According to http://fuse.fusesource.org/mq/docs/mq-fabric.html (Section "Topologies"), configuring ActiveMQ in clustering mode is suitable for non-persistent message. IMHO, it is true that it's not suitable, because all running brokers use the same store file. But, what about separating the store file? Now it's possible right?
Could anybody explain this? If it's not possible, what is the best way to load balance persistent message?
Thanks
You can share the load of persistent messages by creating 2 master/slave pairs. The master and slave share their state either though a database or a shared filesystem so you need to duplicate that setup.
Create 2 master slave pairs, and configure so called "network connectors" between the 2 pairs. This will double your performance without risk of loosing messages.
See http://activemq.apache.org/networks-of-brokers.html
This answer relates to an version of the question before the Camel details were added.
It is not immediately clear what exactly it is that you want to load balance and why. Messages across consumers? Producers across brokers? What sort of concern are you trying to address?
In general you should avoid using networks of brokers unless you are trying to address some sort of geographical use case, have too many connections for a signle broker to handle, or if a single broker (which could be a pair of brokers configured in HA) is not giving you the throughput that you require (in 90% of cases it will).
In a broker network, each node has its own store and passes messages around by way of a mechanism called store-and-forward. Have a read of Understanding broker networks for an explanation of how this works.
ActiveMQ already works as a kind of load balancer by distributing messages evenly in a round-robin fashion among the subscribers on a queue. So if you have 2 subscribers on a queue, and send it a stream of messages A,B,C,D; one subcriber will receive A & C, while the other receives B & D.
If you want to take this a step further and group related messages on a queue so that they are processed consistently by only one subscriber, you should consider Message Groups.
Adding consumers might help to a point (depends on the number of cores/cpus your server has). Adding threads beyond the point your "Camel server" is utilizing all available CPU for the business processing makes no sense and can be conter productive.
Adding more ActiveMQ machines is probably needed. You can use an ActiveMQ "network" to communicate between instances that has separated persistence files. It should be straight forward to add more brokers and put them into a network.
Make sure you performance test along the road to make sure what kind of load the broker can handle and what load the camel processor can handle (if at different machines).
When you do persistent messaging - you likely also want transactions. Make sure you are using them.
If all running brokers use the same store file or tx-supported database for persistence, then only the first broker to start will be active, while others are in standby mode until the first one loses its lock.
If you want to loadbalance your persistence, there were two way that we could try to do:
configure several brokers in network-bridge mode, then send messages
to any one and consumer messages from more than one of them. it can
loadbalance the brokers and loadbalance the persistences.
override the persistenceAdapter and use the database-sharding middleware
(such as tddl:https://github.com/alibaba/tb_tddl) to store the
messages by partitions.
Your first step is to increase the number of workers that are processing from ActiveMQ. The way to do this is to add the ?concurrentConsumers=10 attribute to the starting URI. The default behaviour is that only one thread consumes from that endpoint, leading to a pile up of messages in ActiveMQ. Adding more brokers won't help.
Secondly what you appear to be doing could benefit from a Staged Event-Driven Architecture (SEDA). In a SEDA, processing is broken down into a number of stages which can have different numbers of consumer on them to even out throughput. Your threads consuming from ActiveMQ only do one step of the process, hand off the Exchange to the next phase and go back to pulling messages from the input queue.
You route can therefore be rewritten as 2 smaller routes:
from("activemq:input?concurrentConsumers=10").id("FirstPhase")
.process(businessInvokerOne)
.to("seda:invokeSecondProcess");
from("seda:invokeSecondProcess?concurentConsumers=20").id("SecondPhase")
.process(businessInvokerTwo)
.to("activemq:output");
The two stages can have different numbers of concurrent consumers so that the rate of message consumption from the input queue matches the rate of output. This is useful if one of the invokers is much slower than another.
The seda: endpoint can be replaced with another intermediate activemq: endpoint if you want message persistence.
Finally to increase throughput, you can focus on making the processing itself faster, by profiling the invokers themselves and optimising that code.
I'm setting up a web service with pyramid. A typical request for a view will be very long, about 15 min to finish. So my idea was to queue jobs with celery and a rabbitmq broker.
I would like to know what would be the best way to ensure that bad things cannot happen.
Specifically I would like to prevent the task queue from overflow for example.
A first mesure will be defining quotas per IP, to limit the number of requests a given IP can submit per hour.
However I cannot predict the number of involved IPs, so this cannot solve everything.
I have read that it's not possible to limit the queue size with celery/rabbitmq. I was thinking of retrieving the queue size before pushing a new item into it but I'm not sure if it's a good idea.
I'm not used to good practices in messaging/job scheduling. Is there a recommended way to handle this kind of problems ?
RabbitMQ has flow control built into the QoS. If RabbitMQ cannot handle the publishing rate it will adjust the TCP window size to slow down the publishers. In the event of too many messages being sent to the server it will also overflow to disk. This will allow your consumer to be a bit more naive although if you restart the connection on error and flood the connection you can cause problems.
I've always decided to spend more time making sure the publishers/consumers could work with multiple queue servers instead of trying to make them more intelligent about a single queue server. The benefit is that if you are really overloading a single server you can just add another one (or another pair if using RabbitMQ HA. There is a useful video from Pycon about Messaging at Scale using Celery and RabbitMQ that should be of use.