Perceval retrievers

Perceval is the GrimoireLab component retrieving data from data sources, usually by interacting with service APIs or logs. A part of it is generic for all data sources, but most of its code is composed by retrievers. A retriever is a Perceval component specialized in retrieving data from a specific data source.

Structure of a retriever

A Perceval retriever is built with three components:

  • Client: interacts directly with the data source.
  • Backend: orchestrates the fetching process by using the Client.
  • CommandLine: defines the arguments to initialize and run the Backend from the command line.

Backend and CommandLine extend the abstract classes in They require the definition of several methods:

  • Backend:

    • metadata_category(item) defines the type of the fetched items (e.g., issue, topic, channel, etc.)
    • metadata_id(item) identifies the unique ID of the fetched items (e.g., id of GitHub issues, commit SHA for Git, etc.)
    • metadata_updated_on(item) determines the update time of the fetched items (e.g., updated_at attribute of GitHub issues, committed date for Git, etc.)
    • has_resuming() returns a boolean value whether the backend supports the resuming of the fetch process
    • has_caching() returns a boolean value whether the backend supports caching items during the fetch process.
    • fetch(from_date) contains the logic to perform the fetching process
    • fetch_from_cache() contains the logic to perform the fetching process from cache
  • CommandLine:

    • setup_cmd_parser() initializes the command parser for the backend.

All backends have their own unit tests and corresponding data, saved in the folder /tests and /tests/data respectively. Since most backends fetch data from HTTP APIs, their tests rely on HTTPretty (version==0.8.6), a mocking tool that simulates HTTP requests.

Tips for implementing retrievers

Implementation & conventions:

  • Components are usually implemented incrementally, starting with the Client, Backend, and CommandLine classes.
  • Client methods are named representing the information returned by the backend, using nouns instead of verbs. For example: __issues(..)__ instead of __get_issues(...)__, __user(...)__ instead of __get_user(...)__


Testing is important to ensure that the Perceval retriever is working well:

  • Tests are written together with the components, as soon as possible.
  • Some tests are run on just a tiny subset of the data source, such as a small set of GitHub issues or commits in Git.
  • Once the retriever is near complete, stress tests are run too, such as retrieving data from very large repositories.


The cache is filled with raw items produced during the fetch process, not JSON documents produced by the Backend. Fetching data from cache is probably one of the most tricky activities when implementing a backend. The complexity lays on the fact that the cache, basically a FIFO queue, stores items which in many cases have a nested tree structure. This happens because usually items of different kinds are retrieved from the data source. For example, for an issue tracking system, issues, comments and authors could be retrieved. This nested structure has to be correctly stored, and later retrieved when calling the fetch_from_cache() method.

A strategy to deal with such a complexity is to add markers before and after pushing items to the cache. Examples of this strategy have been implemented for GitHub and Launchpad backends.

The Python-like pseudocode below shows this strategy for extracting issues, comments and their authors.

def fetch():
    raw_issues = client.fetch_issues()

    for issue in json.loads(raw_issues)
        raw_comments = client.fetch_comments(issue)

        for comment in json.loads(raw_comments)
            raw_author = client.fetch_author(comment)

def fetch_from_cache():
    raw_items = cache.retrieve()
    raw_item = next(raw_items)
    while raw_item != '{}{}'

        if raw_item == '{ISSUES}'
            raw_issues = next(raw_items)
            issues = json.loads(raw_issues)

            for issue in issues
                raw_item = next(raw_items)
                while raw_item != '{ISSUE-END}'
                    if raw_item == '{COMMENTS}'
                        raw_comments = next(raw_items)
                        comments = json.loads(raw_comments)
                        for comment in comments
                            tag_author = next(raw_items)
                            raw_author = next(raw_items)

                    raw_item = next(raw_items)
            raw_item = next(raw_items)    

results matching ""

    No results matching ""