If you happen to’re performing internet requests with Spring Boot’s WebClient you maybe, similar to us, learn that defining the URL of your request needs to be executed utilizing a URI builder (e.g. Spring 5 WebClient):
If that’s the case, we suggest that you just ignore what you learn (until looking hard-to-find reminiscence leaks is your pastime) and use the next for establishing a URI as a substitute:
On this weblog publish we’ll clarify how one can keep away from reminiscence leaks with Spring Boot WebClient and why it’s higher to keep away from the previous sample, utilizing our private expertise as motivation.
How did we uncover this reminiscence leak?
Some time again we upgraded our software to make use of the most recent model of the Axle framework. Axle is the bol.com framework for constructing Java functions, like (REST) companies and frontend functions. It closely depends on Spring Boot and this improve additionally concerned updating from Spring Boot model 2.3.12 to model 2.4.11.
When operating our scheduled efficiency assessments, the whole lot regarded high-quality. Most of our software’s endpoints nonetheless offered response instances of beneath 5 milliseconds. Nonetheless, because the efficiency take a look at progressed, we observed our software’s response instances growing as much as 20 milliseconds, and after an extended operating load take a look at over the weekend, issues obtained loads worse. The response instances skyrocketed to seconds – not good.
Earlier than Axle improve: Higher ninetieth percentile response instances of one in all our endpointsAfter Axle improve: Higher ninetieth percentile response instances of the identical endpoint
After an extended stare down contest with our Grafana dashboards, which offer insights into our software’s CPU, thread and reminiscence utilization, this reminiscence utilization sample caught our eye:
This graph exhibits the JVM heap measurement earlier than, throughout, and after a efficiency take a look at that ran from 21:00 to 0:00. In the course of the efficiency take a look at, the applying created threads and objects to deal with all incoming requests. So, the capricious line displaying the reminiscence utilization throughout this era is strictly what we might count on. Nonetheless, when the mud from the efficiency take a look at settles down, we might count on the reminiscence to additionally settle right down to the identical stage as earlier than, however it’s really greater. Does anybody else odor a reminiscence leak?
Time to name within the MAT (Eclipse Reminiscence Analyzer Instrument) to seek out out what causes this reminiscence leak.
What prompted this reminiscence leak?
To troubleshoot this reminiscence leak we:
- Restarted the applying.
- Carried out a heap dump (a snapshot of all of the objects which might be in reminiscence within the JVM at a sure second).
- Triggered a efficiency take a look at.
- Carried out one other heap dump as soon as the take a look at finishes.
This allowed us to make use of MAT’s superior function to detect the leak suspects by evaluating two heap dumps taken a while aside. However we didn’t need to go that far, since, the heap dump from after the take a look at was sufficient for MAT to seek out one thing suspicious:
Right here MAT tells us that one occasion of Spring Boot’s AutoConfiguredCompositeMeterRegistry occupies nearly 500MB, which is 74% of the entire used heap measurement. It additionally tells us that it has a (concurrent) hashmap that’s liable for this. We’re nearly there!
With MAT’s dominator tree function, we are able to record the most important objects and see what they saved alive – That sounds helpful, so let’s use it to have a peek at what’s within this humongous hashmap:
Utilizing the dominator tree we had been capable of simply flick through the hashmap’s contents. Within the above image we opened two hashmap nodes. Right here we see a number of micrometer timers tagged with “v2/merchandise/…” and a product id. Hmm, the place have we seen that earlier than?
What does WebClient need to do with this?
So, it’s Spring Boot’s metrics which might be liable for this reminiscence leak, however what does WebClienthave to do with this? To search out that out you actually have to know what causes Spring’s metrics to retailer all these timers.
Inspecting the implementation of AutoConfiguredCompositeMeterRegistrywe see that it shops the metrics in a hashmap named meterMap. So, let’s put a well-placed breakpoint on the spot the place new entries are added and set off our suspicious name our WebClientperforms to the “v2/product/{productId}” endpoint.
We run the applying once more and … Gotcha! For every name the WebClientmakes to the “v2/product/{productId}” endpoint, we noticed Spring creating a brand new Timerfor every distinctive occasion of product identifier. Every such timer is then saved within the AutoConfiguredCompositeMeterRegistrybean. That explains why we see so many timers with tags like these:
/v2/merchandise/9200000109074941 /v2/merchandise/9200000099621587
How are you going to repair this reminiscence leak?
Earlier than we establish when this reminiscence leak may have an effect on you, let’s first clarify how one would repair it. We’ve talked about within the introduction, that by merely not utilizing a URI builder to assemble WebClient URLs, you’ll be able to keep away from this reminiscence leak. Now we are going to clarify why it really works.
After just a little on-line analysis we got here throughout this publish (https://rieckpil.de/expose-metrics-of-spring-webclient-using-spring-boot-actuator/) of Philip Riecks, by which he explains:
“As we normally need the templated URI string like “/todos/{id}” for reporting and never a number of metrics e.g. “/todos/1337” or “/todos/42″ . The WebClient affords a number of methods to assemble the URI […], which you’ll all use, besides one.”
And that methodology is utilizing the URI builder, coincidentally the one we’re utilizing:
Certainly, once we assemble the URI like that, the reminiscence leak disappears. Additionally, the response instances are again to regular once more.
When may the reminiscence leak have an effect on you? – a easy reply
Do it is advisable fear about this reminiscence leak? Properly, let’s begin with the obvious case. In case your software exposes its HTTP shopper metrics, and makes use of a way that takes a URI builder to set a templated URI onto a WebClient, it is best to positively be nervous.
You possibly can simply examine in case your software exposes http shopper metrics in two alternative ways:
- Inspecting the “/actuator/metrics/http.shopper.requests” endpoint of your Spring Boot software after your software made at the very least one exterior name. A 404 means your software doesn’t expose them.
- Checking if the worth of the applying property administration.metrics.allow.http.shopper.metrics is ready to true, by which case your software does expose them.
Nonetheless, this doesn’t imply that you just’re protected in case you’re not exposing the HTTP shopper metrics. We’ve been passing templated URIs to the WebClient utilizing a builder for ages, and we’ve by no means uncovered our HTTP shopper metrics. But, hastily this reminiscence leak reared its ugly head after an software improve.
So, may this reminiscence leak have an effect on you then? Simply don’t use URI builders together with your WebClient and you need to be protected in opposition to this potential reminiscence leak. That might be the easy reply. You do not take easy solutions? Truthful sufficient, learn on to seek out out what actually prompted this for us.
When may the reminiscence leak have an effect on you? – a extra full reply
So, how did a easy software improve trigger this reminiscence leak to rear its ugly head? Evidently, the addition of a transitive Prometheus (https://prometheus.io/) dependency – an open supply monitoring and alerting framework – prompted the reminiscence leak in our specific case. To grasp why, let’s return to the state of affairs earlier than we added Prometheus.
Earlier than we dragged within the Prometheus library, we pushed our metrics to statsd (https://github.com/statsd/statsd) – a community daemon that listens to and aggregates software metrics despatched over UDP or TCP. The StatsdMeterRegistry that’s a part of the Spring framework is liable for pushing metrics to statsd. The StatsdMeterRegistry solely pushes metrics that aren’t filtered out by a MeterFilter. The administration.metrics.allow.http.shopper.metrics property is an instance of such a MeterFilter. In different phrases, if administration.metrics.allow.http.shopper.metrics = false the StatsdMeterRegistry will not push any HTTP shopper metric to statsd and will not retailer these metrics in reminiscence both. To date, so good.
By including the transitive Prometheus dependency, we added one more meter registry to our software, the PrometheusMeterRegistry. When there’s a couple of meter registry to reveal metrics to, Spring instantiates a CompositeMeterRegistry bean. This bean retains observe of all particular person meter registries, collects all metrics and forwards them to all of the delegates it holds. It’s the addition of this bean that prompted the difficulty.
The problem is that MeterFilter cases aren’t utilized to the CompositeMeterRegistry, however solely to MeterRegistry cases within the CompositeMeterRegistry (See this commit for extra info.) That explains why theAutoConfiguredCompositeMeterRegistryaccumulates all of the HTTP shopper metrics in reminiscence, even once we explicitly set administration.metrics.allow.http.shopper.metricsto false.
Nonetheless confused? No worries, simply don’t use URI builders together with your WebClient and you need to be protected in opposition to this reminiscence leak.
Conclusion
On this weblog publish we defined that this method of defining URLs of your request with Spring Boot’s WebClient is greatest averted:
