Greatly improve performance of the diff parser.
Review Request #674 — Created May 30, 2017 and updated — Latest diff uploaded
Updated by me
The diff parser had some major performance problems that didn't really
affect standard diffs, but were absolutely noticeable for multi-megabyte
The primary performance problems had to do with the building of the diff
data strings. We were concatenating strings, which is not fast or
memory-efficient. To solve this, a breaking change to the diff parser
API had to be made (which is unlikely to affect too many people out
there). We no longer work with the data string, but instead call methods
to prepend or append data to the diff's ParsedDiffFile object (renamed
from File, while we're breaking stuff anyway -- further changes will be
coming in for parsing/attribute/doc improvements). These methods in turn
work on a cStringIO, which is a much faster way of performing the string
Along with this, there were many places where we did conditional lookups
or range checks to determine whether to process one or more lines of
content. These checks were preventing us from unintentionally indexing
past the end of the diff. Since in each case it's more likely than not
that the current line will be somewhere before the end of the diff, the
parser now attempts the index unconditionally, and handles alternative
logic if hitting an IndexError.
Prior to these speedups, a 13MB diff took so long to parse that I gave
up trying to time it. After, it took a few seconds. This should
certainly help lighten the load on busy Review Board servers, and may
even make general uploading of diffs faster in many real-world cases.
All unit tests passed.
Uploaded a very large diff. Saw that it parsed correctly in seconds.
Reviewed at https://reviews.reviewboard.org/r/8880/