Retry partial failures

Hi,

Some AWS APIs can complete with partial success. For example, when calling a PutRecords request with 500 entries, it is possible for some of the records to successfully get put and some fail.

PutRecordsRequest request = new PutRecordsRequest();
// add 500 entries to the request
PutRecordsResult response = producer.putRecords(request);
// response.getFailedRecordCount() could be between 0 and 500
// For example, the first 300 records might get inserts and the last 200 fail due to exceeding throughput 

In such cases:

1. the client should wait a bit,
2. retrieve the index of failed records from response and fetch those failed records to build a new request
3. re-call producer.putRecords with the new request

So every invocation needs access to the result of the previous execution (or null for the first execution).

Is this possible to do with failsafe? From what I can see invocations do not have access to results of previous attempts.

[BenRomberg]

First of all, great library - we use it for over a year now and are quite happy with it.

We have the exact same use-case (PutRecords for AWS Kinesis) and have worked around Failsafe’s limitations like this:

AtomicReference<List<PutRecordsRequestEntry>> currentRequestEntries = new AtomicReference<>(putRecordsRequestEntries);
List<PutRecordsRequestEntry> unsuccessfulRequestEntries = Failsafe.with(
        new RetryPolicy().<List<PutRecordsRequestEntry>>retryIf(
                recordsToRetry -> !recordsToRetry.isEmpty()))
        .get(() -> {
            List<PutRecordsRequestEntry> newRequestEntries = putRecords(currentRequestEntries.get());
            currentRequestEntries.set(newRequestEntries);
            return newRequestEntries;
        });

Would be great if Failsafe could support this so we can remove the AtomicReference workaround.

[jhalterman]

With the addition of ExecutionContext.getLastResult and getLastFailure I’m going to consider this closed. Feel free to comment and/or reopen if you feel those are inadequate for your use case.