Concurrency limiter from Netflix to avoid throttling storage #2169
Conversation
…lector is either Kafka or RabbitMQ
There was a problem hiding this comment.
thanks for giving this a go.
first question: have you tried this successfully?
second comment: I think the real constraint is storage so this should be a property at global level not nested for each collector. otherwise your constraints won't apply if both http and Kafka spans are accepted.
we can work through other design things after consideration of above 2 things.
|
Hi Adrian, Do you have any design concern how shall concurrency limiter applied on http collector? The http collector is a special case, because http client knows synchronously whether collection succeeded or not, and current ConcurrentLimiter works as a pool of threads, processing requests asynchronously. |
|
Hi again, Regards, |
|
yeap on this being where I was hoping it would be (will apply to all collectors, not just http but good start) once you have your results in, we can chat further. thanks so far. |
|
@malonso1976 were you able to try this out in a real environment? |
|
Not really, but I am on it.
I stopped working on it because I had a family emergency that took me away
from work a couple of months.
I am back and want to test it soon.
El El mié, 24 oct 2018 a las 2:11, Tommy Ludwig <[email protected]>
escribió:
… @malonso1976 <https://github.com/malonso1976> were you able to try this
out in a real environment?
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#2169 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AFx3OsKlvOke_5BR9OqMV-CEbwJztqLIks5un7A_gaJpZM4WCgvm>
.
|
|
@malonso1976 No problem at all. Let us know when you have a chance to test it out. |
| @@ -64,6 +64,7 @@ | |||
| <!-- Be careful to set this as a provided dep, so that it doesn't interfere with dependencies | |||
| from other projects. For example, cassandra and spring boot set guava versions --> | |||
| <guava.version>19.0</guava.version> | |||
| <netflix.concurrency.limits.version>0.0.49</netflix.concurrency.limits.version> | |||
There was a problem hiding this comment.
@adriancole, concurrency-limits v0.2.0 is the latest tag but uses Java 8 features (such as Optional). Do you know of any current plans to get io.zipkin.zipkin2:zipkin off of 1.6 and up to at least 1.8? I saw #777 but it doesn't appear to have actually upped the main.signature.artifact.
It also looks like retrolambda does not support recompiling dependencies or I'd explore that option further.
There was a problem hiding this comment.
Never mind, made a new artifact outside of zipkin-core so it's not subject to this limitation. Affords me some other luxuries around not-shading dependencies as well :)
|
zipkin-collector is java 8 for this reason. we can shade some smaller deps
if needed
…On Wed, Apr 10, 2019, 7:53 AM Logic-32 ***@***.***> wrote:
***@***.**** commented on this pull request.
------------------------------
In pom.xml
<#2169 (comment)>:
> @@ -64,6 +64,7 @@
<!-- Be careful to set this as a provided dep, so that it doesn't interfere with dependencies
from other projects. For example, cassandra and spring boot set guava versions -->
<guava.version>19.0</guava.version>
+ <netflix.concurrency.limits.version>0.0.49</netflix.concurrency.limits.version>
Never mind, made a new artifact outside of zipkin-core so it's not subject
to this limitation. Affords me some other luxuries around not-shading
dependencies as well :)
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#2169 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AAD611-pgR227k6QeCjFbcjaT_sdWxVJks5vfRn8gaJpZM4WCgvm>
.
|
Adding storage-throttle module/etc. to contain logic for wrapping other storage implementations and limiting the number of requests that can go through to them at a given time. Elasticsearch storage's maxRequests can be override by throttle properties if the throttle is enabled. Making sure RejectedExecutionExceptions are "first class" citizens since they are used to reduce the throttle. Removing HttpCall's Semaphore in favor of the throttle (same purpose, different implementations). Inspired by work done on openzipkin#2169.
|
#2502 will address this use case. Thanks |
Adding storage-throttle module/etc. to contain logic for wrapping other storage implementations and limiting the number of requests that can go through to them at a given time. Elasticsearch storage's maxRequests can be override by throttle properties if the throttle is enabled. Inspired by work done on openzipkin#2169.
Adding ThrottledStorageComponent/etc. to contain logic for wrapping other storage implementations and limiting the number of requests that can go through to them at a given time. Elasticsearch storage's maxRequests can be override by throttle properties if the throttle is enabled. Inspired by work done on openzipkin#2169.
Adding ThrottledStorageComponent/etc. to contain logic for wrapping other storage implementations and limiting the number of requests that can go through to them at a given time. Elasticsearch storage's maxRequests can be override by throttle properties if the throttle is enabled. Inspired by work done on openzipkin#2169.
|
#2502 now includes test instructions. please give a try! |
Adding ThrottledStorageComponent/etc. to contain logic for wrapping other storage implementations and limiting the number of requests that can go through to them at a given time. Elasticsearch storage's maxRequests can be override by throttle properties if the throttle is enabled. Inspired by work done on openzipkin#2169.
Adding ThrottledStorageComponent/etc. to contain logic for wrapping other storage implementations and limiting the number of requests that can go through to them at a given time. Elasticsearch storage's maxRequests can be override by throttle properties if the throttle is enabled. Inspired by work done on #2169.
…nzipkin#2502) Adding ThrottledStorageComponent/etc. to contain logic for wrapping other storage implementations and limiting the number of requests that can go through to them at a given time. Elasticsearch storage's maxRequests can be override by throttle properties if the throttle is enabled. Inspired by work done on openzipkin#2169.
|
note: concurrency-limits-core has been abandoned by Netflix by over a year at this point. We may have to revisit this |
|
I opened this noticing no release since last july even if there were a couple more months of changes, nothing was released Netflix/concurrency-limits#157 |
|
I've heard the word resilience4j lately - never used it but just a reference |
|
Reslience4j and Failsafe are interesting options but, when evaluating them, make sure they solve the problem this is looking to address. Namely, storage gets backed up so we need to slow down how much we're writing to it but do so in a way that doesn't cause the server to exhaust memory. In this particular case, a limited-size queue is used to buffer some data while we have a small bottleneck we're working with. Some simple retries with an appropriate backoff policy could address the same basic problem of slow storage but may be harder to tame with respect to pending actions. Meaning, I can't say how hard it is to prevent an "infinite queue" with the retry handlers. FWIW, the ExecutorService that backs this implementation is pretty universal and supported by Java. The only part that is not is the one that adjusts how many threads are allowed to be active in it. To a certain extent, that logic could be directly incorporated into Zipkin instead of pulled in via Netflix. |
|
@Logic-32 Thanks for summarizing the problem, it'll help if we look into any migration! I was indeed incorrectly thinking the https://github.com/resilience4j/resilience4j#632-threadpoolbulkhead |
No description provided.