I have a situation where I host a high RPS highly available service that receives requests aka commands. These commands have to be sent to N downstream clients, who actually execute them. Each downstream client is separate microsevice and has different constraints like mode (sync,async), execution cadence etc.
Should a slow downstream client build the logic to receive all requests and execute them in batches as they want ? Or my service should build logic to talk to slow and fast clients by maintaining state for commands across downstream clients. Share your opinions
Not enough info to give any prescriptive advice, but I'd start with dividing the tasks into async and sync first. Those are 2 completely different workloads that, most likely, would require different implementation stacks. I'll give you an idea of what you can start with in the world of AWS...
Not knowing what you mean by async, I'd default to a message-bus setup. In that case you can use something like Amazon Kinesis or Kafka for ingestion purposes, and kicking off Lambda or EC2 instance. If the clients need to be notified of a finished job they can either long-poll an SQS queue, subscribe to an SNS topic, or use MQTT with websockets for a long-running connection.
The sync tasks are easier, since it's all about processing power. Just make sure you have your EC2 instances in an auto-scaling group behind an ALB or API Gateway to scale out, and in, appropriately.
This is a very simple answer since I don't have any details needed to be more precise, but this should give you an idea of where to get started.
Related
We have a bus reservation system running in GKE in which we are handling the creation of such reservations with different threads. Due to that, CRUD java methods can sometimes run simultaneously referring to the same bus, resulting in the save in our DB of the LAST simultaneous update only (so the other simultaneous updates are lost).
Even if the probabilities are low (the simultaneous updates need to be really close, 1-2 seconds), we need to avoid this. My question is about how to address the solution:
Lock the bus object and return error to the other simultaneous requests
In-memory map or Redis caché to track the bus requests
Use GCP Pub/Sub, Kafka or RabbitMQ as a queue system.
Try to focus the efforts on reducing the simultaneous time window (reduce from 1-2 seconds up to milliseconds)
Others?
Also, we are worried if in the future the GKE requests handling scalability may be an issue. If we manage a relatively higher number of buses, should we need to implement a queue system between the client and the server? Or GKE load balancer & ambassador will already manages it for us? In case we need a queue system in the future, could it be used also for the collision problem we are facing now?
Last, the reservation requests from the client often takes a while. Therefore, we are changing the requests to be handled asynchronously with a long polling approach from the client to know the task status. Could we link this solution to the current problem? For example, using the Redis caché or the queue system to know the task status? Or should we try to keep the requests synchronous and focus on reducing the processing time (it may be quite difficult).
PROBLEM
Our PROCESSING SERVICE is serving UI, API, and internal clients and listening for commands from Kafka.
Few API clients might create a lot of generation tasks (one task is N messages) in a short time. With Kafka, we can't control commands distribution, because each command comes to the partition which is consumed by one processing instance (aka worker). Thus, UI requests could be waiting too long while API requests are processing.
In an ideal implementation, we should handle all tasks evenly, regardless of its size. The capacity of the processing service is distributed among all active tasks. And even if the cluster is heavily loaded, we always understand that the new task that has arrived will be able to start processing almost immediately, at least before the processing of all other tasks ends.
SOLUTION
Instead, we want an architecture that looks more like the following diagram, where we have separate queues per combination of customer and endpoint. This architecture gives us much better isolation, as well as the ability to dynamically adjust throughput on a per-customer basis.
On the side of the producer
the task comes from the client
immediately create a queue for this task
send all messages to this queue
On the side of the consumer
in one process, you constantly update the list of queues
in other processes, you follow this list and consume for example 1 message from each queue
scale consumers
QUESTION
Is there any common solution to such a problem? Using RabbitMQ or any other tooling. Нistorically, we use Kafka on the project, so if there is any approach using - it is amazing, but we can use any technology for the solution.
Why not use spark to execute the messages within the task? What I'm thinking is that each worker creates a spark context that then parallelizes the messages. The function that is mapped can be based on which kafka topic the user is consuming. I suspect however your queues might have tasks that contained a mixture of messages, UI, API calls, etc. This will result in a more complex mapping function. If you're not using a standalone cluster and are using YARN or something similar you can change the queueing method that the spark master is using.
As I understood the problem, you want to create request isolation from the customer using dynamically allocated queues which will allow each customer tasks to be executed independently. The problem looks like similar to Head of line blocking issue in networking
The dynamically allocating queues is difficult. This can also lead to explosion of number of queues that can be a burden to the infrastructure. Also, some queues could be empty or very less load. RabbitMQ won't help here, it is a queue with different protocol than kafka.
One alternative is to use custom partitioner in kafka that can look at the partition load and based on that load balance the tasks. This works if the tasks are independent in nature and there is no state store maintains in the worker.
The other alternative would be to load balance at the customer level. In this case you select a dedicated set of predefined queues for a set of customers. Customers with certain Ids will be getting served by a set of queues. The downside of this is some queues can have less load than others. This solution is similar to Virtual Output Queuing in networking,
My understanding is that the partitioning of the messages it's not ensuring a evenly load-balance. I think that you should avoid create overengineering and so some custom stuff that will come on top of the Kafka partitioner and instead think at a good partitioning key that will allows you to use Kafka in an efficiently manner.
A little background.
Very big monolithic Django application. All components use the same database. We need to separate services so we can independently upgrade some parts of the system without affecting the rest.
We use RabbitMQ as a broker to Celery.
Right now we have two options:
HTTP Services using a REST interface.
JSONRPC over AMQP to a event loop service
My team is leaning towards HTTP because that's what they are familiar with but I think the advantages of using RPC over AMQP far outweigh it.
AMQP provides us with the capabilities to easily add in load balancing, and high availability, with guaranteed message deliveries.
Whereas with HTTP we have to create client HTTP wrappers to work with the REST interfaces, we have to put in a load balancer and set up that infrastructure in order to have HA etc.
With AMQP I can just spawn another instance of the service, it will connect to the same queue as the other instances and bam, HA and load balancing.
Am I missing something with my thoughts on AMQP?
At first,
REST, RPC - architecture patterns, AMQP - wire-level and HTTP - application protocol which run on top of TCP/IP
AMQP is a specific protocol when HTTP - general-purpose protocol, thus, HTTP has damn high overhead comparing to AMQP
AMQP nature is asynchronous where HTTP nature is synchronous
both REST and RPC use data serialization, which format is up to you and it depends of infrastructure. If you are using python everywhere I think you can use python native serialization - pickle which should be faster than JSON or any other formats.
both HTTP+REST and AMQP+RPC can run in heterogeneous and/or distributed environment
So if you are choosing what to use: HTTP+REST or AMQP+RPC, the answer is really subject of infrastructure complexity and resource usage. Without any specific requirements both solution will work fine, but i would rather make some abstraction to be able switch between them transparently.
You told that your team familiar with HTTP but not with AMQP. If development time is an important time you got an answer.
If you want to build HA infrastructure with minimal complexity I guess AMQP protocol is what you want.
I had an experience with both of them and advantages of RESTful services are:
they well-mapped on web interface
people are familiar with them
easy to debug (due to general purpose of HTTP)
easy provide API to third-party services.
Advantages of AMQP-based solution:
damn fast
flexible
cost-effective (in resources usage meaning)
Note, that you can provide RESTful API to third-party services on top of your AMQP-based API while REST is not a protocol but rather paradigm, but you should think about it building your AQMP RPC api. I have done it in this way to provide API to external third-party services and provide access to API on those part of infrastructure which run on old codebase or where it is not possible to add AMQP support.
If I am right your question is about how to better organize communication between different parts of your software, not how to provide an API to end-users.
If you have a high-load project RabbitMQ is damn good piece of software and you can easily add any number of workers which run on different machines. Also it has mirroring and clustering out of the box. And one more thing, RabbitMQ is build on top of Erlang OTP, which is high-reliable,stable platform ... (bla-bla-bla), it is good not only for marketing but for engineers too. I had an issue with RabbitMQ only once when nginx logs took all disc space on the same partition where RabbitMQ run.
UPD (May 2018):
Saurabh Bhoomkar posted a link to the MQ vs. HTTP article written by Arnold Shoon on June 7th, 2012, here's a copy of it:
I was going through my old files and came across my notes on MQ and thought I’d share some reasons to use MQ vs. HTTP:
If your consumer processes at a fixed rate (i.e. can’t handle floods to the HTTP server [bursts]) then using MQ provides the flexibility for the service to buffer the other requests vs. bogging it down.
Time independent processing and messaging exchange patterns — if the thread is performing a fire-and-forget, then MQ is better suited for that pattern vs. HTTP.
Long-lived processes are better suited for MQ as you can send a request and have a seperate thread listening for responses (note WS-Addressing allows HTTP to process in this manner but requires both endpoints to support that capability).
Loose coupling where one process can continue to do work even if the other process is not available vs. HTTP having to retry.
Request prioritization where more important messages can jump to the front of the queue.
XA transactions – MQ is fully XA compliant – HTTP is not.
Fault tolerance – MQ messages survive server or network failures – HTTP does not.
MQ provides for ‘assured’ delivery of messages once and only once, http does not.
MQ provides the ability to do message segmentation and message grouping for large messages – HTTP does not have that ability as it treats each transaction seperately.
MQ provides a pub/sub interface where-as HTTP is point-to-point.
UPD (Dec 2018):
As noticed by #Kevin in comments below, it's questionable that RabbitMQ scales better then RESTful servies. My original answer was based on simply adding more workers, which is just a part of scaling and as long as single AMQP broker capacity not exceeded, it is true, though after that it requires more advanced techniques like Highly Available (Mirrored) Queues which makes both HTTP and AMQP-based services have some non-trivial complexity to scale at infrastructure level.
After careful thinking I also removed that maintaining AMQP broker (RabbitMQ) is simpler than any HTTP server: original answer was written in Jun 2013 and a lot of changed since that time, but the main change was that I get more insight in both of approaches, so the best I can say now that "your mileage may vary".
Also note, that comparing both HTTP and AMQP is apple to oranges to some extent, so please, do not interpret this answer as the ultimate guidance to base your decision on but rather take it as one of sources or as a reference for your further researches to find out what exact solution will match your particular case.
The irony of the solution OP had to accept is, AMQP or other MQ solutions are often used to insulate callers from the inherent unreliability of HTTP-only services -- to provide some level of timeout & retry logic and message persistence so the caller doesn't have to implement its own HTTP insulation code. A very thin HTTP gateway or adapter layer over a reliable AMQP core, with option to go straight to AMQP using a more reliable client protocol like JSONRPC would often be the best solution for this scenario.
Your thoughts on AMQP are spot on!
Furthermore, since you are transitioning from a monolithic to a more distributed architecture, then adopting AMQP for communication between the services is more ideal for your use case. Here is why…
Communication via a REST interface and by extension HTTP is synchronous in nature — this synchronous nature of HTTP makes it a not-so-great option as the pattern of communication in a distributed architecture like the one you talk about. Why?
Imagine you have two services, service A and service B in that your Django application that communicate via REST API calls. This API calls usually play out this way: service A makes an http request to service B, waits idly for the response, and only proceeds to the next task after getting a response from service B. In essence, service A is blocked until it receives a response from service B.
This is problematic because one of the goals with microservices is to build small autonomous services that would always be available even if one or more services are down– No single point of failure. The fact that service A connects directly to service B and in fact, waits for some response, introduces a level of coupling that detracts from the intended autonomy of each service.
AMQP on the other hand is asynchronous in nature — this asynchronous nature of AMQP makes it great for use in your scenario and other like it.
If you go down the AMQP route, instead of service A making requests to service B directly, you can introduce an AMQP based MQ between these two services. Service A will add requests to the Message Queue. Service B then picks up the request and processes it at its own pace.
This approach decouples the two services and, by extension, makes them autonomous. This is true because:
If service B fails unexpectedly, service A will keep accepting requests and adding them to the queue as though nothing happened. The requests would always be in the queue for service B to process them when it’s back online.
If service A experiences a spike in traffic, service B won’t even notice because it only picks up requests from the Message Queues at its own pace
This approach also has the added benefit of being easy to scale— you can add more queues or create copies of service B to process more requests.
Lastly, service A does not have to wait for a response from service B, the end users don’t also have to wait for long— this leads to improved performance and, by extension, a better user experience.
Just in case you are considering moving from HTTP to AMQP in your distributed architecture and you are just not sure how to go about it, you can checkout this 7 parts beginner guide on message queues and microservices. It shows you how to use a message queue in a distributed architecture by walking you through a demo project.
I'm fairly new to Akka and writing concurrent applications and I'm wondering what's a good way to implement an actor that would wait for a redis list and once an item becomes available it will process it, or send it to a different actor to process?
Would using the blocking function BRPOPLPUSH be better, or would a scheduler that will ask the actor to poll redis every second be a better way?
Also, on a normal system, how many of these actors can I spawn concurrently without consuming all the resource the system has to offer? How does one decide how many of each Actor type should an actor system be able to handle on the system its running on?
As a rule of thumb you should never block inside receive. Each actor should rely only on CPU and never wait, sleep or block on I/O. When these conditions are met you can create even millions of actors working concurrently. Each actor is suppose to have 600-650 bytes memory footprint (see: Concurrency, Scalability & Fault-tolerance 2.0 with Akka Actors & STM).
Back to your main question. Unfortunately there is no official Redis client "compatible" with Akka philosophy, that is, completely asynchronous. What you need is a client that instead of blocking will return you a Future object of some sort and allow you to register callback when results are available. There are such clients e.g. for Perl and node.js.
However I found fyrie-redis independent project which you might find useful. If you are bound to synchronous client, the best you can do is either:
poll Redis periodically without blocking and inform some actor by sending a message to with a Redis reply or
block inside an actor and understand the consequences
See also
Redis client library recommendations for use from Scala
BRPOPLPUSH with block for long time (up to the timeout you specify), so I would favour a Scheduler instead which still blocks, but for a shorter amount of time every second or so.
Whichever way you go, because you are blocking, you should read this section of the Akka docs which describes methods for working with blocking libraries.
Do you you have control over the code that is inserting the item into redis? If so you could get that code to send your akka code a message (maybe over ActiveMQ using the akka camel support) to notify it when the item has been inserted into redis. This will be a more event driven way of working and prevent you from having to poll, or block for super long periods of time.
Suppose we were to implement a network application, such as a chat with a central server and several clients: we assume that all communication must go through the central server, then it should pick up messages from some clients and forward them to target clients, and so on.
Regardless of the technology used (sockets, web services, etc..), it is possible to think that there are some producer threads (that generate messages) and some consumer threads (that read messages).
For example, you could use a single queue for incoming and outgoing messages, but using a single queue, you couldn't receive and send messages simultaneously, because only one thread at a time can access the queue.
Perhaps it would be more appropriate to use two queues: for example, this article explains a way in which you can manage a double queue so that producers and consumers can work almost simultaneously. This scenario may be fine if there are only a producer and a consumer, but if there are many clients:
How to make so that the central server can receive data simultaneously from multiple input streams?
How to make so that the central server can send data simultaneously to multiple output streams?
To resolve this problem, my idea is to use a double queue for each client: on the central server, each client connection may be associated with two queues, one for incoming messages from that client and one for outgoing messages addressed to that client. In this way the central server may send and receive data simultaneously on almost all the connections with the clients...
There are probably other ways to manage the queues ... What are the parameters to determine how many queues are needed and how to organize them? There are cases that do not need any queue?
To me, this idea of using a queue per client or multiple queues per client seems to miss the point. First of all, it is absolutely possible to build a queue which can be accessed simultaneously by 2 threads (one can be enqueueing an item while a different one is dequeueing another item). If you want to know how, post a specific question about that.
Second, even if we assume that only 1 thread at a time can access a single queue, and even if we assume that the server will be receiving or sending data to/from all the clients simultaneously, it still doesn't follow that you need a different queue for each client. To avoid limiting system performance, you just need to allow enough concurrency to utilize all the server's CPUs. Even with a single, system-wide queue, if dequeueing/enqueueing messages is fast enough compared to the other work the server is doing, it might not be a bottleneck. (And with an efficient implementation, simply inserting an item or removing an item from a queue should be very fast. It's a very simple operation.) For that message queue to become the bottleneck limiting performance, either you would need a LOT of CPUs, or everything else the server was doing would have to be very fast. In that case, you could work out some scheme with 2 or 4 system-wide queues, to allow 2x or 4x more concurrency.
The whole idea of using work queues in a multi-threaded system is that they 1) allow multiple consumers to all grab work from a single location, so producers can "dump" whatever work they need done at that single location without worrying about which consumer will do it, and 2) function as a load-balancing mechanism for the consumers. (Additionally, a work queue can act as a "buffer" if producers temporarily generate work too fast for the consumers.) If you have a dedicated pair of producer-consumer threads for each client, it calls into question why you need to use queues at all. Why not just do a synchronous "pass off" from dedicated producer to corresponding dedicated consumer? Or, why not use a single thread per client which acts as both producer and consumer? Using queues in the way which you are proposing doesn't seem to really gain anything.