[PIP-149] Making the REST Admin API fully async

Co-author: @mattisonchao @Technoboy-

Motivation

The Rest API was originally designed to be implemented asynchronously, but with the iteration of functions, some synchronous implementations were added, resulting in many asynchronous methods called synchronous implementations. Also, many synchronous calls do not add timeouts. This greatly reduces concurrency, user operations, and experience. In order to prevent more problems, and improve code readability and maintainability, we intend to refactor these synchronous calls and standardize the implementation of the API.

Related discussion: https://lists.apache.org/thread/pkkz2jgwtzpksp6d4rdm1pyxzb3z6vmg

Goals

• Try to avoid synchronous method calls in asynchronous methods.
• Async variable (AsyncResponse) is placed in the first parameter position.
• Async variable (AsyncResponse) cannot be substituted into method implementations.
• Add more tests and increase the coverage.

Modification

1. Avoid synchronous method calls in asynchronous methods.

   protected void internalDeleteNamespace(boolean authoritative) {
      validateTenantOperation(namespaceName.getTenant(), TenantOperation.DELETE_NAMESPACE);
      validatePoliciesReadOnlyAccess();
   }
   

   Suggest to do like this:

   protected CompletableFuture<Void> internalDeleteNamespace(boolean authoritative) {
       return validateTenantOperationAsync(namespaceName.getTenant(), TenantOperation.DELETE_NAMESPACE)
      .thenCompose(__ -> validatePoliciesReadOnlyAccessAsync());
   }
   

2. Async variable (AsyncResponse) is placed in the first parameter position

   public void deleteNamespace(@Suspended final AsyncResponse asyncResponse, 
            @PathParam("tenant") String tenant,
            @PathParam("namespace") String namespace,
            @QueryParam("force") @DefaultValue("false") boolean force,
            @QueryParam("authoritative") @DefaultValue("false") boolean authoritative) {
   
   

3. Async variable (AsyncResponse) cannot be substituted into method implementations

   internalCreateNonPartitionedTopicAsync(asyncResponse, authoritative, properties);
   

   Suggest to do like this:

   internalCreateNonPartitionedTopicAsync(authoritative, properties)
           .thenAccept(__ -> asyncResponse.resume(Response.noContent().build()))
           .exceptionally(ex -> {
               resumeAsyncResponseExceptionally(asyncResponse, ex.getCause());
               return null;
           });
   

4. Exception Some methods will validate ownership, like namespace ownership, topic ownership, so will throw REDIRECT exception. we need to filter this exception and not print log.

 internalCreateNonPartitionedTopicAsync(authoritative, properties)
        .thenAccept(__ -> asyncResponse.resume(Response.noContent().build()))
        .exceptionally(ex -> {
            if (!isRedirectException(ex)) {
                 log.error("Failed to xxx");
            }
            resumeAsyncResponseExceptionally(asyncResponse, ex.getCause());
            return null;
        });

Task tracking

In order to unify the modification and track the modified part, it’s better to open an issue to track, like Apache/Pulsar#14353, Apache/Pulsar#14013, Apache/Pulsar#13854.

[lhotari]

@codelipenghui @lhotari @eolivelli @BewareMyPower @nodece @mattisonchao I have opened this issue to discuss first then decide to make to PIP

Thank you @Technoboy- . Good work. I shared my thoughts and some expectations in yesterdays Pulsar Community meeting, you can find the discussion and @merlimat’s responses of the rationale for the changes. It would be good to document the arguments that Matteo explained why mixing async and blocking (sync) causes issues. Please check the last 35 minutes from the recording of the meeting (it’s not available yet, @merlimat could you please add the recording on https://github.com/apache/pulsar/wiki/Community-Meetings#recordings ?) I’ll be able to follow up at the beginning of March the next time. There’s no need to wait for me to reply.

While digging into the Jetty settings in Pulsar, I noticed a few gaps in backpressure handling, which are relevant when there are more requests which are handled asynchronously. I have a draft PR #14353 . I’ll resume work on that in March. The current values for queue sizes and thread pool sizes are just guesses. Most likely we will use much lower values to prevent the broker taking in too much work in parallel. That’s the essence of back pressure that it limits the in progress work so that incoming requests also slow down. Currently that is not the case since the thread pool queue can grow in an unbounded way (LinkedBlockingQueue is used under the covers). There are several kludges that attempt to add backpressure, but they aren’t very effective in Pulsar currently. #14353 will help address backpressure issues in Pulsar Admin API. These problems will come more evident when there are more APIs which are implemented using asynchronous Servlet API. /cc @merlimat @codelipenghui

[Technoboy-]

@codelipenghui @lhotari @eolivelli @BewareMyPower @nodece @mattisonchao I have opened this issue to discuss first then decide to make to PIP