diff

znkr.io/diff

A high-performance difference algorithm module for Go.

Difference algorithms compare two inputs and find the edits that transform one to the other. This is very useful to understand changes, for example when comparing a test result with the expected result or to understand which changes have been made to a file.

This module provides diffing for arbitrary Go slices and text.

I wrote a bit about the background and the design decisions that went into this module on flo.znkr.io/diff.

Installation

To use this module in your Go project, run:

go get znkr.io/diff

API Documentation

Full documentation available at pkg.go.dev/znkr.io/diff.

Examples

Comparing Slices

Diffing two slices produces either the full list of edits

x := strings.Fields("calm seas reflect the sky")
y := strings.Fields("restless seas reflect the sky defiantly")
edits := diff.Edits(x, y)
for i, edit := range edits {
    if i > 0 {
        fmt.Print(" ")
    }
    switch edit.Op {
    case diff.Match:
        fmt.Printf("%s", edit.X)
    case diff.Delete:
        fmt.Printf("[-%s-]", edit.X)
    case diff.Insert:
        fmt.Printf("{+%s+}", edit.Y)
    default:
        panic("never reached")
    }
}
// Output:
// [-calm-] {+restless+} seas reflect the sky {+defiantly+}

or a list of hunks representing consecutive edits

x := strings.Fields("calm seas reflect the sky")
y := strings.Fields("restless seas reflect the sky defiantly")
hunks := diff.Hunks(x, y, diff.Context(1))
for i, h := range hunks {
    if i > 0 {
        fmt.Print(" … ")
    }
    for i, edit := range h.Edits {
        if i > 0 {
            fmt.Print(" ")
        }
        switch edit.Op {
        case diff.Match:
            fmt.Printf("%s", edit.X)
        case diff.Delete:
            fmt.Printf("[-%s-]", edit.X)
        case diff.Insert:
            fmt.Printf("{+%s+}", edit.Y)
        default:
            panic("never reached")
        }
    }
}
// Output:
// [-calm-] {+restless+} seas … sky {+defiantly+}

For both functions, a ...Func variant exists that works with arbitrary slices by taking an equality function, and a ...Hash variant that takes a maphash.Hasher. Use the ...Hash variant instead of the ...Func variant if the elements can be hashed. It's about 25% slower than Hunks and Edits, so use those for comparable types.

Comparing Text

Because of its importance, comparing text line by line has special support and produces output in the unified diff format:

x := `this paragraph
is not
changed and
barely long
enough to
create a
new hunk

this paragraph
is going to be
removed
`

y := `this is a new paragraph
that is inserted at the top

this paragraph
is not
changed and
barely long
enough to
create a
new hunk
`
fmt.Print(textdiff.Unified(x, y))
// Output:
// @@ -1,3 +1,6 @@
// +this is a new paragraph
// +that is inserted at the top
// +
//  this paragraph
//  is not
//  changed and
// @@ -5,7 +8,3 @@
//  enough to
//  create a
//  new hunk
// -
// -this paragraph
// -is going to be
// -removed

Stability

API: Stable - The API is stable and follows the Go module compatibility guidelines.

The exact diff output is not guaranteed to be stable: performance and quality improvements will likely change the output of a diff. Committing to a stable diff result would be too limiting.

Diff Readability

Diffs produced by this module are intended to be readable by humans.

Readable diffs have been the subject of a lot of discussions and have even resulted in some new diffing algorithms like the patience or histogram algorithms in git. However, the best work about diff readability by far is diff-slider-tools by Michael Haggerty. He implemented a heuristic that's applied in a post-processing step to improve the readability. This module implements this heuristic in the textdiff package.

For example:

x := `// ...
["foo", "bar", "baz"].map do |i|
  i.upcase
end
`

y := `// ...
["foo", "bar", "baz"].map do |i|
  i
end

["foo", "bar", "baz"].map do |i|
  i.upcase
end
`

fmt.Println("With textdiff.IndentHeuristic:")
fmt.Print(textdiff.Unified(x, y, textdiff.IndentHeuristic()))
fmt.Println()
fmt.Println("Without textdiff.IndentHeuristic:")
fmt.Print(textdiff.Unified(x, y))
// Output:
// With textdiff.IndentHeuristic:
// @@ -1,4 +1,8 @@
//  // ...
// +["foo", "bar", "baz"].map do |i|
// +  i
// +end
// +
//  ["foo", "bar", "baz"].map do |i|
//    i.upcase
//  end
//
// Without textdiff.IndentHeuristic:
// @@ -1,4 +1,8 @@
//  // ...
//  ["foo", "bar", "baz"].map do |i|
// +  i
// +end
// +
// +["foo", "bar", "baz"].map do |i|
//    i.upcase
//  end

Performance

By default, the underlying diff algorithm used is Myers' algorithm augmented by a number of heuristics to speed up the algorithm in exchange for non-minimal diffs. The diff.Minimal option is provided to skip these heuristics to get a minimal diff independent of the costs and diff.Fast to use a fast heuristic to get a non-minimal diff as fast as possible.

On an M1 Mac, the default settings almost always result in runtimes < 1 ms, but truly large diffs (e.g. caused by changing generators for generated files) can result in runtimes of about 100 ms. Below is the distribution of runtimes applying textdiff.Unified to every commit in the Go repository (y-axis is in log scale). It's produced by go run ./internal/cmd/eval -repo <repo> -stats stats.csv -validate=false and the notebook in plots:

Comparison with other Implementations

Comparing the performance with other Go modules that implement the same features is always interesting, because it can surface missed optimization opportunities. This is especially interesting for larger inputs where superlinear growth can become a problem. Below are benchmarks of znkr.io/diff against other popular Go diff modules:

• znkr: Default configuration with performance optimizations enabled
• znkr-minimal: With diff.Minimal() option for minimal diffs
• znkr-fast: With diff.Fast() option for fastest possible diffing
• go-internal: Patience diff algorithm from github.com/rogpeppe/go-internal
• diffmatchpatch: Implementation from github.com/sergi/go-diff
• godebug: Implementation from github.com/kylelemons/godebug
• mb0: Implementation from github.com/mb0/diff
• udiff: Implementation from github.com/aymanbagabas/go-udiff

Note: It's possible that the benchmark is using diffmatchpatch incorrectly, the benchmark numbers certainly look suspiciously high. However, the way it's used in the benchmark is used in at least one large open source project.

Runtime Performance (seconds per operation)

On the benchmarks used for this comparison znkr.io/diff almost always outperforms the other implementations. However, there's one case where go-internal and udiff are significantly faster, but the resulting diffs are 10% larger (see numbers below). On that case, diff.Fast is faster than both and produces diffs of the same size. The numbers are from an M1 Pro, produced with the command in internal/benchmarks.

Diff Minimality (number of edits produced)

Correctness

I tested this diff implementation against every commit in the Go repository using the standard unix patch tool to ensure that all diff results are correct.

This test is part of the test suite for this module and can be run with

go run ./internal/cmd/eval -repo <repo>

License

This module is distributed under the Apache License, Version 2.0, see LICENSE for more information.