spring-cloud-netflix
spring-cloud-netflix copied to clipboard
Connection/Read Timeouts with Load Balanced Rest Template
We are migrating our load balanced RestTemplates from using the legacy, netflix http client to the new RestTemplate + Spring Retry.
One thing that surprised us was that the ribbon setting for connection timeouts are NOT applied when using the rest template. This actually seems like an obvious fact once you think about it, but it means that timeouts for the load balanced rest templates really need to be configured differently then how you configure those same timeouts in zuul (and likely Feign).
The connection/read timeouts for a load balanced rest template are NOT configured via the ribbon properties. Instead, you must set your timeouts when constructing the rest template via setConnectionTimeout
and setReadTimeout
in the restTemplateBuilder.
We ended up normalizing our code such that it just injects those same ribbon timeouts settings into our template:
@Value("${ribbon.ConnectionTimeout:1000}")
int connectionTimeout;
@Value("${ribbon.ReadTimeout:20000}")
int readTimeout;
@Bean
@LoadBalanced
public RestTemplate buildRestTemplateRibbon(RestTemplateBuilder restTemplateBuilder) {
return restTemplateBuilder.messageConverters(new BuildHttpMessageConverters().getConverters())
.uriTemplateHandler(new BuildUriTemplateHandler())
.setConnectTimeout(connectionTimeout)
.setReadTimeout(readTimeout)
.errorHandler(new ServiceClientErrorHandler()).build();
}
It would be nice if that were the default behavior in spring cloud, as then the timeouts are set the same across zuul, feign, and the load balanced rest template.
Or alternatively, it might be good to make a note in the documentation: If you are using the load balanced rest template with retry, you need to explicitly set your read/connection timeouts when constructing the template.
I don't think most people were using the legacy RestClient
+ RestTemplate
. I think documentation is the way to go especially since we're not in the business of creating RestTemplate
objects.
Just to clarify, we are no longer using the legacy RestClient, we are moving to the new and improved retry logic in Camden.SR6.
I just wanted to highlight that with the new spring retry logic in Camden.SR6 and Dalston:
The connection timeouts for Zuul and Feign still rely on the ribbon.ConnectionTimeout and ribbon.ReadTimeout. However, the Load Balanced rest templates do not use those two settings.
This becomes important when doing retries because connection timeout and read timeout exceptions are "retryable". If you fail to set the read timeouts on your rest template, your rest template's retries will work differently than your zuul/Feign timeouts.
I think we understand what you are saying. I think an argument for the opposite can also be made, when using a RestTemplate
you may expect that it would be configured using the normal RestTemplate
configuration. In other words, Ribbon configuration would not have any effect.
Documentation is definitely warranted.
Makes sense to me, appreciate the quick response.
NP, we appreciate the feedback!
If Documentation is definitely warranted, what would it say?
I'm somewhat lost on how to find/set and understand the different socket timing issues. Documentation would be helpful.
@DaveBender
If you are using Zuul or the Feign client you can use the
ribbon.ConnectionTimeout ribbon.ReadTimeout
If you are using a RestTemplate with the @LoadBalanced
annotations, you need to set your socket timeouts via the RestTemplateBuilder :
@Bean
@LoadBalanced
public RestTemplate restTemplateRibbon(RestTemplateBuilder restTemplateBuilder) {
return restTemplateBuilder.
.setConnectTimeout(connectionTimeout)
.setReadTimeout(readTimeout)
.build();
If you end up wrapping your RestTemplate calls within a Hystrix circuit breaker, you will want to set the hystrix timeout HIGHER than the RestTemplate's timeouts:
Hystrix Timeout > (connect timeout + read timeout) * (Number of ribbon retries)
A cautionary tale: If you create a RestTemplate via its default constructor (not using a builder to set socket timeouts), it can potentially wait indefinitely because timeouts are NOT defined.
@ryanjbaxter We need to set connection timeout and read time out specific to each load balancing client(microservice). So in this case, if we configure connect timeout and read timeout direclty to RestTemplate, it would contribute to all clients right ? How to configure timeouts per client(microservice) ?
I think in that case you would have to create a RestTemplate
object per service and not one global RestTemplate
used everywhere.
@geowalrus4gh We are doing what Ryan suggested, we create different RestTemplates for our various third-party clients that have different timeout requirements. One thing we do to control this process is we configure a RestTemplateBuilder with some sensible defaults:
@Bean
public RestTemplateBuilder restTemplateBuilder() {
RestTemplateBuilder builder = new RestTemplateBuilder();
HttpMessageConverters converters = this.messageConverters.getIfUnique();
if (converters != null) {
builder = builder.messageConverters(converters.getConverters());
}
List<RestTemplateCustomizer> customizers = this.restTemplateCustomizers
.getIfAvailable();
if (!CollectionUtils.isEmpty(customizers)) {
customizers = new ArrayList<RestTemplateCustomizer>(customizers);
AnnotationAwareOrderComparator.sort(customizers);
builder = builder.customizers(customizers);
}
builder = builder.setConnectTimeout(connectionTimeout);
builder = builder.setReadTimeout(readTimeout);
return builder;
}
And then when we need to override those defaults, we do so in a PostConstruct method:
@Autowired
private RestTemplateBuilder restTemplateBuilder;
protected RestTemplate restTemplate;
@PostConstruct
public void init() {
restTemplate = restTemplateBuilder
.setConnectTimeout(overrideConnectionTimeout)
.setReadTimeout(overrideReadTimeout)
.messageConverters(new StringHttpMessageConverter())
.build();
}