Skip to content

logdelta 0.3.0

Choose a tag to compare

@antonsoo antonsoo released this 02 Oct 06:09
· 65 commits to main since this release
cargo install logdelta

Checked against real CI logs for the first time: a failing pytest job of the pytest project on GitHub Actions (854 lines) and three passing runs of the same job (about 1,270 lines each), then a vitest job and a tokio job the same way. The numbers below are 0.2.3 from crates.io and this release on the same files.

Added

  • Blocks. Findings whose lines sit together are reported as one. The pytest failure was 311 findings, 543 lines of report: every line of the traceback its own NEW finding, every line of the upload steps the job never reached its own GONE finding. It is now 4: the traceback (51 templates), the short test summary (27), the lines the job exited on (5), and one GONE block of 198 templates. vitest: 26 findings to 6. tokio: 99 to 13.
    • A NEW block is a run of new lines, with up to two known lines let through, found while the target is read. It prints one line per template with its line number, and ×N where a template has more lines. The same stack trace logged 25 times is one block that says so.
    • A GONE block is located in the first baseline, the run that still had the lines.
    • A template folds into a block only if all of its lines are in such groups, so a message that also occurs on its own keeps its own finding and its count. Grouping never makes a report longer.
    • --block-lines N (default 12) sets how much of a long block is printed: its start and its end. --flat turns grouping off.
    • --json gains a blocks array and a block index on each finding that is part of one; findings still lists every template. Memory stays flat: a few numbers per template, nothing per line.
  • A GONE finding shows the line the first baseline had, with its line number (first_baseline_line_no and first_baseline_raw in --json), where it used to show the template alone.
  • Test-runner progress lines are masked: ....s....x.... [ 42%] becomes <PROGRESS> [ <NUM>%], or <PROGRESS!> when it holds an F or E. Which tests land on which line depends on scheduling, so two passing runs of the pytest job diffed to 27 findings, all of them dots. Now 0.
  • OSC escape sequences (hyperlinks, window titles) are stripped before mining, as CSI color codes already were.
  • Library: blocks::Block, DiffResult::blocks, DiffOptions::group, scoring::count_g_test, Drain::add_token_slice, mask::strip_escapes, output::printable.

Changed

  • CHANGED needs the count to move, not only the share. The score compares a template's share of the log, and a share moves whenever the run is shorter or longer for other reasons: the failing pytest job printed its 257 [new branch] lines 258 times in a log a third shorter, and that was a CHANGED finding ("more often", score 38). A second G-test on the counts alone (the baselines' total against the target's, as two Poisson counts) now has to clear the same cutoff, in the same direction. A bigger run with the same mix still changes nothing, which is what the share test is for.
  • A NEW VALUE on a line inside a NEW block is part of the block and is not listed again.
  • Reports print log lines without the log's own escape sequences and control characters. A colored CI log put its color codes into the terminal report (76 lines of the pytest report carried them, also under --color never), where a line cut to the terminal width left the color on for everything after it, and into Markdown tables as stray [31m. --json keeps every line as it was read.
  • The header counts what the report shows and says what --flat would: 4 findings (284 with --flat).
  • The web demo shows the engine's blocks (it grouped new lines on its own before, by first line only), has a GONE block card, cuts a long block to its start and end with a button for the rest, and strips color codes from the lines it shows.

Fixed

  • A line that is mostly placeholders never matched its own template. The similarity between a line and a template skipped positions where both hold the same masking placeholder, but still divided by the full length, so <NUM> <NUM> <HEX> ok scored 1/4 against itself and every such line started a new template. 100,000 lines of that shape were 100,000 templates, 77.55 s and 45.7 MB in templates (each line compared against every template before it); they are one template in 0.18 s and 7.2 MB. In a diff, a line identical in every run was reported NEW: the toolchain line - 1.98.1 aarch64-apple-darwin 48a229ce… in tokio's logs, between two passing runs. Such positions are now left out of both sides of the fraction.