Upload a patch to fix the problem, and add test cases to verify the retry is working properly

1. all raised exceptions need to include the URL of the timeline server in them. Otherwise nobody will ever be able to track down the problem if its any
2. you can actually test TimelineClient (or any Yarn service) in a try-with-resources clause, to get the service automatically stopped {{ Service extends Closeable}}, see

try(Timeline client = createClient()) {
}

Uploaded a new patch.

all raised exceptions need to include the URL of the timeline server in them. Otherwise nobody will ever be able to track down the problem if its any

Make sense because in v2 the url can be changed. Log the url if exception happens when connecting the timeline server.

you can actually test TimelineClient (or any Yarn service) in a try-with-resources clause, to get the service automatically stopped

Done

In addition, I fixed another bug of retry: When we use up retry quota, we should throw IOException (which is declared in the method signature) other than RuntimeException, such that users can catch and process the exception.

Make sense because in v2 the url can be changed. Log the url if exception happens when connecting the timeline server.

makes sense in v1 as people can misconfigure their URL.

Oddly enough, I'm deep in trying to fix some timeline problems, and was about to add better diagnostics myself. I'd like to see this patch (or variant thereof) in branch-2

I'm linking this patch to YARN-3477, in which the TL client doesn't rethrow the underlying exception on a retry timeout. It should, as that preserves the most information.

1. line 407: if e.getCause()==IOE, cast and throw rather than wrap-and-throw

makes sense in v1 as people can misconfigure their URL.

In v1, when we initialize the client, it will print the url, which won't change during the lifecycle of the client.

I'd like to see this patch (or variant thereof) in branch-2

This patch mainly focus on reusing TimelineClientConnectionRetry for v2 APIs to only retry on connect exception, while it fix some related problems of TimelineClientConnectionRetry.

But I agree that we can have a jira to fix TimelineClientConnectionRetry generally on trunk/branch-2.

Looking at this patch, it doesn't address the issues I've encountered in YARN-3477

1. when the retries time out, the exception causing the attempts to fail is not rethrown
2. interrupts are being swallowed, making it impossible to reliably interrupt the thread

I'd like to get the YARN-3477 patches in as well as these ones, so lets see if we can do them in turn.

Varun Saxena Are you working on this? If not, I'd like to take it.

Assigning this to myself, per offline discussion in weekly call

Both 1) and 2) have been resolved after YARN-4675. (TimelineClientConnectionRetry is called in a Jersey client filter that only retries requests upon connection-related exceptions) Closing this. Feel free to reopen it if you disagree.