Rerun Spring integration tests after minimizations with full context reset

[IlyaMuravjov]

Description

Right now we don't fully reset Spring context between concrete executions when generating integration tests with fuzzer, because it's too time consuming (can take several seconds per reset). We do our best to reset relevant parts of context (e.g. reset relevant beans and rollback transactions), however that may still not be enough because, for example, database id generators are not rollbacked with the transaction.

Partial reset of Spring context may lead to generation of unreproducible tests that rely on some code from earlier concrete executions being executed before them.

Action plan

To cope we that, it is suggested to rerun tests that are left after test case minimization with full context reset and use results obtained from these reruns.

[alisevych]

use results obtained from these reruns

@IlyaMuravjov Do you mean to:

• ignore results of the first runs on Dirty context?
• generate tests based on the reruns only?

Also can you define more exactly: what data from results will be used? Is it used for assertions only?

[IlyaMuravjov]

@alisevych Yes, I mean exactly what you said (tests are generated based on results from reruns only, original results obtained using dirty context are ignored).

The data from results includes UtExecution.stateAfter, UtExecution.result, and UtExecution.coverage, they are used for generating assertions, printing stack traces, and generating SARIF reports, and possibly something else.

[alisevych]

I wanted to point out that executions are made twice. That is time consuming.
If run results can be compared and there are the differences, we could make some conclusions from them. You've suggested to just lose this information.
cc @EgorkaKulikov

[IlyaMuravjov]

@alisevych

First, I'd like to pay more attention to the fact that rerun on clean context differs a lot from the original run:

• rerun is much slower (from preliminary results I've noticed that is over an order magnitude slower when testing on the spring-boot-testing project);
• rerun is expected to produce more (or at least not less) reliable results.

If we see the difference in results from runs on clean and dirty context, the only conclusion that we seem to be able to make is that either we have failed to properly reset all relevant parts of context when doing partial reset or the result is indeed non deterministic.

I think, we should only notify the user about non deterministic results after making at least two runs on clean context and observing different results.

Finally, all of the handling differences in run results depends on code generation being ready to deal with multiple unequal results.

[alisevych]

@IlyaMuravjov Thank you! Enjoyed reading.