streamly/CONTRIBUTING.md
2021-01-01 14:13:35 +05:30

10 KiB

Contributors' Guide

Bug Reports

Please feel free to open an issue for any questions, suggestions, issues or bugs in the library. In case you are reporting a bug we will appreciate if you provide as much detail as possible to reproduce or understand the problem, but nevertheless you are encouraged to open an issue for any problem that you may encounter.

Picking Issues to Work on

Beginners are encouraged to pick up issues that are marked help wanted. It is a good idea to update the issue expressing your intent so that others do not duplicate the effort and people with a background on the issue can help.

Contributing A Change

If the feature makes significant changes to design, we encourage you to open an issue as early as possible so that you do not have to redo much work because of changes in design decisions. However, if you are confident, you can still go ahead and take that risk as the maintainers are supposed to be reasonable people.

Pull Requests (PR)

Please feel free to send a pull request (PR) whether it is a single letter typo fix or a complex change. We will accept any PR that makes a net positive change to the package. We encourage you to provide a complete, consistent change with test, documentation and benchmarks. You can contact the maintainers for any help or collaboration needed in that regard. However, if due to lack of time you are not able to complete the PR, you are still welcome to submit it, maintainers will try their best to actively contribute and pick up your change as long as the change is approved.

Pull Request (PR) Checklist

Here is a quick checklist for a PR, for details please see the next section:

  • PR contains one logical changeset
  • Each commit in the PR consists of a logical change
  • Commits are rebased/squashed/fixup/reordered as needed
  • Stylistic changes to irrelevant parts of the code are separated in independent commits.
  • Code is formatted as per the style of the file or that of other files
  • Compiler warnings are fixed
  • Reasonable hlint suggestions are accepted
  • Tests are added to cover the changed parts
  • All tests pass
  • Performance benchmarks are added, where applicable
  • No significant regressions are reported by performance benchmarks
  • Haddock documentation is added to user visible APIs and data types
  • Tutorial module, README, and guides are updated if necessary.
  • Changelog is updated if needed
  • The code conforms to the license, it is not stolen, credit is given, copyright notices are retained, original license is included if needed.

Structuring Pull Requests

Lean PR: Please keep the reviewer in mind when sending pull requests. Use one PR for each logical changeset. A lean and thin PR with one independent logical change is more likely to be reviewed and merged quickly than a big monolithic PR with several unrelated changes.

PR Dependencies: If your change depends on an earlier PR you can create a branch from the old PR's branch, and raise a new PR targeting to merge into the old PR branch. Do not push the commits to the same PR just because the commit depends on that PR.

Commits: You are encouraged to group a logically related set of changes into a single commit. When the overall changeset is largish you can divide it into multiple smaller commits, with each commit having a logically grouped changeset and the title summarizing the logical grouping. Always keep in mind a logical division helps reviewers understand and review your code quickly, easier history tracking and when required clean reversal changes.

Functional Changes: Keep the reviewer in mind when making changes. Please resist the temptation to make style related changes to surrounding code unless you are changing that code anyway . Whenever possible, try to separate unrelated refactoring changes which do not affect functionality in separate commits so that it is easier for the reviewer to verify functional changes.

Style Changes: Make sure that your IDE/Editor is not automatically making sweeping style changes to all the files you are editing. That makes separating the signal from the noise very difficult and makes everything harder. If you would like to make style related changes then please send a separate PR with just those changes and no other functionality changes.

Rebasing: If your commits reflect how you fixed intermediate problems during testing or made random changes at different times you may have to squash your changes (git rebase -i) into a single commit or logically related set of commits.

Resolving Conflicts

If during the course of development or before you send the PR you find that your changes are conflicting with the master branch then use git rebase master to rebase your changes on top of master. DO NOT MERGE MASTER INTO YOUR BRANCH.

Testing

It is a good idea to include tests for the changes where applicable. See the existing tests here.

Documentation

For user visible APIs, it is a good idea to provide haddock documentation that is easily understood by the end programmer and does not sound highfalutin, and preferably with examples. If your change affects the tutorial or needs to be mentioned in the tutorial then please update the tutorial. Check if the additional guides are affected or need to updated.

Performance Benchmarks

It is a good idea to run performance benchmarks to see if your change affects any of the existing performance benchmarks. If you introduced something new then you may want to add benchmarks to check if it performs as well as expected by the programmers to deem it usable.

See the README file in the benchmark directory for more details on how to run the benchmarks.

The first level of check is the regression in "allocations", this is the most stable measure of regression. "bytesCopied" is another stable measure, it gives an idea of the amount of long lived data being copied across generations of GC.

The next measure to look at is "cpuTime" this may have some variance from run to run because of factors like cpu frequency scaling, load on the system or the number of context switches etc. However, the variance would usually be within 5-10%, anything more than that is likely to be a red flag.

Quite often an increase in "cpuTime" would correspond to an increase in "allocations". Increase in "allocations" indicates an inefficiency in computing due to too much GC activity e.g. due to lack of fusion. However, it is also possible for the cpu time to go up without the allocations going up, this indicates an inefficiency in processing in general or more CPU being used for the same task. It could be because of inefficient code generation, branch mis-predictions or lack of cache locality etc.

For concurrent benchmarks we can compare "cpuTime" and "time" to check the degree of concurrency and total efficiency.

Changelog

Any new changes that affect the user of the library in some way must be documented under Unreleased section at the top of the Changelog. The changes in the changelog must be organized in the following categories, in that order:

  • Breaking Changes
  • Enhancements
  • Bug Fixes
  • Deprecations

If there are very few changes then you can just prefix a bullet with these annotations. If there are many changes make sections to group them. A section can be absent if there is nothing to add in it.

If you make changes that are incompatible with the released versions of the library please indicate that in the Changelog as Breaking Changes and also write short notes regarding what the programmers need to do to adapt their existing code to the new change.

Licensing

If you have copied code from elsewhere you need to conform to the licensing terms of the original code. Usually you would need to add a copyright notice in the source header, and include the license of the original code as per the terms of the license. If the code you are copying does not have an associated license please do not use it.

Developer documentation

To build haddock documentation for all modules, including the ones not exposed to users of the library, use the following commands:

$ cabal haddock
$ stack haddock --no-haddock-deps

For design documentation see the design directory.

Coding Guidelines

Coding Style

Please see the Haskell coding style guide.

StreamD coding style

Some conventions that we follow in the StreamD code are illustrated by the following example:

mapM f (Stream step1 state1) = Stream step state
    where

    step gst st = do
        r <- step1 (adaptState gst) st
        case r of
            Yield x s -> f x >>= \a -> return $ Yield a s
            Skip s    -> return $ Skip s
            Stop      -> return Stop
  • For the input streams use numbering for step and state e.g. step1/state1. For the output stream use step and state.
  • For state argument of step, use st.
  • For result of executing a step use r or res
  • For the yielded element use x
  • For the yielded state use s

In general, the rule is - the shorter the scope of a variable the shorter its name can be. For example, s has the shortest scope in the above code, st has a bigger scope and state has the biggest scope.

Design guides

See the design directory for design guidelines and documents.

Tricky Parts

The state-passing through each API is currently fragile. Every time we run a stream we need to be careful about the state we are passing to it. In case of folds where there is no incoming state, we start with the initial state defState. When we have an incoming state passed to us there are two cases:

  1. When we are building a concurrent stream that needs to share the same SVar we pass the incoming state as is.
  2. In all other cases we must not share the SVar and every time we pass on the state to run a stream we must use adaptState to reset the SVar in the state.

When in doubt just use adaptState on the state before passing it on, we will at most lose concurrency but the behavior will be correct.

There is no type level enforcement about this as of now, and therefore we need to be careful when coding. There are specific tests to detect and report any problems due to this, all transform operations must be added to those tests.