chop-dbhi / avocado / 691

80%

Pull #306

travis-ci

Refactor formatter and exporter APIs to reduce overhead

A call to formatter(...) was doing a lot of work for every row being
processed. The primary slow spots are:

- Format methods being dynamically accessed via getattr() on every loop
- Unpacking keyword arguments (**kwargs) in the method call
- Building and moving OrderedDict values
- Logging a message when a value failed be coerced

To remediate these issues, targeted format methods are now processed
at initialization rather that process time. This requires the list of
formats to be passed into the constructor and not in the __call__ method.

Keyword arguments passed in during a call are no longer unpacked in the
method call, but are passed as a keyword argument itself, `kwargs`.

Use of OrderedDict has been removed in favor of a lightweight namedtuple
class as input (for named values) and plain tuples for output. Value
fields are now defined in the `get_*_header` methods. A custom method
can be implemented per format type.

Exceptions that occur during formatting are no long silently skipped
unless the `ExpectedFormatterException` is used.

The exporters have been updated to accommodate this changes. In addition,
new read methods have been implemented for performance testing including
cached and threaded versions. The previous read method have been renamed
to manual_read since it performs manual limit and offsetting and distinct
checking.

To accommodate the new read method, the `write` on all exporters have been
changed to not call `read` itself. This change requires the passed iterable
to already contain the formatted rows.

Note: This refactor requires backwards incompatible changes to get the
performance gains.

Signed-off-by: Byron Ruth <b@devel.io>

204 of 204 new or added lines in 9 files covered. (100.0%)

3172 of 3893 relevant lines covered (81.48%)

3.26 hits per line