@cacoco is there a suggested way to do Request level global processing/error handling? I tried the following (it has some type error atm) but it seems like an overkill. Finch had a RequestMonad that had this idea, not sure about finatra. The goal is that we have all these domain objects that will throw very similar errors like NotFound, DublicateKey etc. so I am curious if we can have all the controllers call one function in stead of repeatedly doing try catch blocks for every end point
commonRequestHandler[T, R, M <: {def toResponse(entity: T): Response}](f: () => T, request: R, mapper: M): Response = {
try{
responseMapper.toResponse(callback())
}catch{
// Catch all the standard errors here and convert to corresponding response
}
}
@sinanspd did you take a look at the user guide section on exception mapping which goes into detail about this? https://twitter.github.io/finatra/user-guide/http/exceptions.html
The key is the Exception Mapping filter which translates things in the response path of the server that is on the way out and back to the caller
This filter if it encounters an exception looks up in the exception manager if there is a mapper that will translate the exception into a prope me httpnl reponse
You would never need to use a mapper directly
So all you need to do is throw proper exceptions and you can define a way for the server to translate those exceptions to responses.
Hope that helps
Hello! I was wondering if there are any general recommendations for configuring a Finagle Client + Finatra Server combination to ensure that the client doesn't overwhelm the server. In our use case, the client is part of a scheduled job that sends 100s of thousands of requests, however the server needs to be able to respond to other users as well. The server was originally Spring-based but seems to be causing more problems now that we've converted it to Finatra.
We've made the connection pool smaller on the client side, but this is resulting in service creation failures after the job has been running for some time. We expected connection requests to be queued rather than seeing these errors - is there anything obvious we might be missing?
Hi there, would need more information (like what errors you're seeing if any or the pattern of behavior exhibited or specifics about what your server metrics looked like) but have you looked at the Finagle client and server documentation? Also, are you blocking the main thread by any chance? https://twitter.github.io/finagle/guide/Clients.html and https://twitter.github.io/finagle/guide/Servers.html and https://twitter.github.io/finagle/guide/Metrics.html and https://finagle.github.io/blog/2016/09/01/block-party/
Hi Christopher, thanks for the reply! Yes we've gone through the docs, and are using Futures all the way through. The specific error we're seeing is
exceeded 2.seconds to <service> while creating a service/connection or reserving a service/connection from the service/connection pool in the Finagle client.
feel free to play around with the params
@sinanspd I would never recommend to do a
Thread.sleep in tests. @genesis-root running a “load test” on a single machine locally is likely to not really tell you anything. You most likely want to perform benchmarking (microbenchmarking) with something like JMH. For true load testing you’d want a replica of your production environment and load generated most likely from many machines to simulate more than what a single box can generate.
@agutteridge can you share how you create and configure your client?
I would also take a look at the client pooling documentation: https://twitter.github.io/finagle/guide/Clients.html#pooling
@genesis-root if you really want something you can run “locally”, the apache bench (ab) tool can be useful: http://httpd.apache.org/docs/current/programs/ab.html
StaticPartitioning currently requires flag 'kafka.application.server' to be set to the Aurora service proxy hostname e.g. 0.tweet-word-count.prod.team1.service.smf1.twitter.com:12345
Can this be used with out Aurora?
@pobrienAvlino To use FinatraKafkaStream's StaticPartitioning trait without Aurora, you can set "kafka.application.server" to be "<instance_id>.:<port>". Make sure the <instance_id> matches the value in the "kafka.current.shard" flag. The <port> isn't used but must be included. The port can be any value greater than 0. For example, to represent shard 5, you could specify kafka.application.server=5:1
@jdcohen220 , you can use the process flag of "-kafka.consumer.auto.offset.reset=latest", which means to start from latest if there is no initial offset. If you want to start from the latest with existing offset, you can use the Stream Reset tool, https://cwiki.apache.org/confluence/display/KAFKA/KIP-171+-+Extend+Consumer+Group+Reset+Offset+for+Stream+Application
Hi guys - I'm having a lot of issues with HttpClient failing and never recovering. I feel like I must be doing something completely wrong - I was assuming that I make one Http.client and then re-use it for multiple requests - is that right? (I'm getting a ChannelClosedException but I'm not seeing any HTTP traffic on the server it is supposed to be hitting)
We've tried setting a bunch of different FailFast and FailureAccrual settings but these don't seem to have helped us
Is there a new example of using the QueryableState? I have gotten through my initial exceptions, but the examples here https://twitter.github.io/finatra/user-guide/kafka-streams/index.html#queryable-state and here https://twitter.github.io/finatra/user-guide/kafka-streams/examples.html#queryable-state don't match the source code I am looking through. While I am able to create the streams correctly and as far as I can tell the stores are created as well. The stores show up in the topology, and I can see them in the admin page. However when I try to access these stores via queryableFinatraKeyValueStore, the stores are not found. I can't even find the stores in FinatraStoresGlobalManager.getStores ... I was unable to locate in the source code where these stores would be put into the FinatraStoresGlobalManager object.
@pobrienAvlino Our apologies that the QueryableState example isn't up to date! The Finatra QueryableState trait currently can only be used when using a FinatraTransformer (which adds some functionality on top of Kafka's Transformer class). The example shown here: https://twitter.github.io/finatra/user-guide/kafka-streams/examples.html#queryable-state is using ".count()(Materialized.as(countStoreName))" which means a normal Kafka's transformer is being used. A correct example would have used FinatraDslWindowedAggregations#sum instead
@pobrienAvlino To implement "Queryable State" with the built in Kafka aggregations and state stores, you can use the KafkaStreams class as such e.g.
Note: Typing the following on my phone so won't likely compile :-)
final override def configureThrift(router: ThriftRouter): Unit = {
router
.add(
new WordCountQueryService(
kafkaStreams
.store(countStoreName)
.asInstanceOf[ReadOnlyKeyValueStore[String, Long]]
)
)
}class WordCountQueryService(
//Note: Make sure ReadOnlyKeyValueStore is "pass by name"
store: => ReadOnlyKeyValueStore[String, Long]]) {
...
}
Hello, has anyone had this issue? I have these case classes
case class Item(priority: Double)
case class Response(serverTime: Long, timeZoneOffset: Int, data: Item)
so the response of my endpoint looks like this
{
"serverTime": "1565292753751",
"timeZoneOffset": "-18000000",
"data": {
"priority": "18.3333"
}
}why the priority field is a string even when I declared as a double?
FinatraJacksonModule made the conversion
I thought the MDC logging was supposed to transparently move across threads via Contexts... what am I missing do I need to configure something for the value to flow outside of the controlled context?
@Singleton
class MyController @Inject()(myService: MyService) extends Controller {
post("/test") { rq: MyRequest =>
info("controller")
myService.test(rq)
}
}
will give me:
@martinffx if you log from an injected class, that needs to be happening within the scope of an initialized MDC context (as described here: https://twitter.github.io/finatra/user-guide/logging/mdc.html#usage). If you are calling the MDC in the injected class this still needs to happen in the scope of an outer initialization. If you are doing that and it is not working, it would be great to file an issue.
I'm also having issues with the Client, it is constantly marking itself as dead and not coming back in a timely manner.
is there a recommended config for low volume servers?
@martinffx there are recommendations if your client to is pointing so a single host (like a VIP). You would disable failfast and failure accrual. See "circuit breaking" here: https://twitter.github.io/finagle/guide/Clients.html