From Fedora Project Wiki

Revision as of 13:22, 20 April 2011 by Jlaska (talk | contribs) (Adjust headings for better section separation ... dunno if this is good, it just reads easier)

Warning.png
This page is a draft only
It is still under construction and content may change. Do not rely on the information on this page.

What is ResultsDB

ResultsDB is a database, in which we'll store results of tests executed by the AutoQA framework.

Motivation behind ResultsDB

At the moment, all the results are stored on the autoqa-results mailing list [1]. Although this is not really a bad system for storing the results, searching for some particular result (even as simple query as "All results for foobar-1.3-fc15") is not really comfortable. ResultsDB's main purpose is not to store the data, but to provide easy access to the results via simple querying API.

This means, that using the ResultsDB as database backend, many 'frontends' can be created. Be it simple tools which will just aggregate most recent results for specified package (maybe a table inside koji/bodhi info pages), some statistical tool for gathering fail/pass ratio of *insert test name of your choice* for given Fedora release, etc.

Current state

From development point of view, there is quite well-defined database schema [2], 'input' API [3], and proof-of-concept TurboGears2 application [4], implementing these two.

I have also started to work on a proof-of-concept frontend [5], which will allow us to create test plans (like Package Update Acceptance Testplan [6]), using mediawiki pages to define the testplan's requirements [7] [8] (this will be covered in more detail further in the text).

API

We have quite well-defined 'input' API - i.e. the API for storing data into ResultsDB. The 'output' API needs to be designed. I have not gone to that particular area that much, because at the time being. My plan was to sneak ResultsDB reporting into the production to gain some reasonable dataset, and then adjust the output API based on the actual needs of our frontend(s).

Note: we all decided against 'sql-based' API, and prefer the set of specific filter-methods with monitorable (?) arguments.

Also, once Fedora Message Bus is up'n'running (if ever), we definitely want to send 'announces' via the bus too.


Thoughts behind the database schema

It's safe to ignore the 'Job' idea at the moment - the idea behind Jobs was that we wanted to be able to internally group several test-runs (e.g. rpmgguard & conflicts for some particular package) into the job. This option is a) IMHO not really feasible with the current way of scheduling tests b) easily substituted with the 'generic testplan frontend' (see below).

Other than that - our primary goal was versatility - we wanted to be able to store results not only from the tests we did have at the moment, but to be able to meet the needs of 'almost anything'.

'Testrun' table corresponds with a single execution of a test, all the data not covered by it's attributes (such as tested package/update, architecture, etc) are to be stored in the 'TestrunData' as key-value pairs. Metadata for each test should define a set of used keyval pairs (with possible information about 'always present' and 'sometimes present' keys) - e.g. [9].

Testruns can be devided into phases (e.g. Setup phase of the test, downloading packages from koji...) - hence the TestrunPhase table. If the test developer decides to split the test into multiple phases, each phase can have it's result. Also metadata can be used to specify how to treat failures/warnings/etc from each phase - warning in setup phase might be less important than warning in run-once, etc.

Generic Testplan frontend

The idea behind this tool is quite simple - Model-View-Controller [10] architecture. ResultsDB will provide data (model), metadata on Mediawiki [8] would be Controller, and the simple-logic frontend [5] is the View part. Given that we probably will want to have multiple test plans, this architecture might suit us quite well.

The [6] is great example. There are multiple 'levels' of acceptance, based on the results of different tests. Our multi-purpose test are run on most (if not all) of the packages anyway, so why not just use the data, and present it in different contexts?

Our TCMS is MediaWiki at the moment, so wwoods proposed quite elegant way of storing (and retrieving, of course) metadata on some wiki page [7]. I just used that idea to store a 'more complicated' kind of template [8].

The semantics of metadata on [8] is quite simple. Let's take this particular

  • testcases (required): A dictionary, which connects name of the testcase (key), and the url of testcase metadata (value). Only testcase names (i.e. rpmlint, initscripts, etc) are used from this point further.
  • testcase_classes (required): A list, which defines any number of 'classes'. These are to be user-defined further in the metadata.
  • mandatory, introspection, advisory (test classes, user defined): Each test class is a dictionary which describes which testcases are to be taken into account, which results will be accepted, and what result should the test class have, if either of the tests is out of the specified set.
    • testcases: List of testcase names to be used in this testclass
    • pass: List of 'accepted' results. If all of the test results are in this set, overall test class result is "PASSED".
    • on_fail: If any of the test results is not in the 'pass' list, the overall test class result is the value defined here.

The frontend [5] then simply parses the metadata, queries ResultsDB for test results for given NVR/testcase combination, and displays it.

Links