What it does

Five ways to run Rift. Every output on this page ran on files made by rift-gen --seed 7, and each box says where it shows selected lines of a longer output.

Plain diff and --summary unified diff, then the counts

rift a b writes a unified diff that patch applies to the first file to get the second. --summary counts first, so you know how much differs before you write a diff.

$ rift a.sql b.sql
--- a.sql+++ b.sql@@ -380,7 +380,6 @@ INSERT INTO public.orders (id, customer_id, status, amount, note) VA… 381	178722	cancelled	2930.54	wmojdqnzik INSERT INTO public.orders (id, customer_id, status, amount, note) VA…-383	668107	shipped	5325.74	zu INSERT INTO public.orders (id, customer_id, status, amount, note) VA… 385	359182	cancelled	475.15	ztbyhdlmxk INSERT INTO public.orders (id, customer_id, status, amount, note) VA…

$ rift --summary a.sql b.sql | grep -E "^(a:|b:|equal|differs|peak heap|time)"
a: 843633 lines, 67108879 bytes
b: 845034 lines, 67182379 bytes
equal    436 segments, 835632 lines in a, 835632 in b
differs  435 segments, 8001 lines in a, 9402 in b
peak heap 78.6 MiB of 256.0 MiB (82397510 bytes)
time 0.17 s

Guide: plain diff, --summary.

Key mode rows compared by key, not by position

An export written in a different row order diffs by content, not by position. --key takes a CSV column, an NDJSON field or a pg_dump COPY column, and lists the keys that were added (+), removed (-) or changed (~), with the line of the record in each file.

$ rift --key csv:id a.csv b.csv > keys.txt
$ for k in "~" "-" "+"; do grep -m 2 "^$k" keys.txt; done
$ cut -c1 keys.txt | sort | uniq -c
~6650	6419	6634
~6651	6420	6635
-144	145	-
-145	146	-
+191	-	140
+192	-	141
4024 -
6535 +
5047 ~

Guide: key mode. Use case: compare CSV files by key.

Git rift as the diff driver

Git calls any file over 512 MiB binary and prints no diff. Name the large files in .gitattributes and git runs rift git-diff for them, with git's own headers. It ships in 0.3.0.

$ echo '*.sql diff=rift' >> .gitattributes
$ git config diff.rift.command 'rift git-diff --mem 1G'
$ git diff dump.sql > out.diff
$ head -n 10 out.diff | cut -c1-90
diff --git a/dump.sql b/dump.sqlindex 86d4681..0000000 100644--- a/dump.sql+++ b/dump.sql@@ -380,7 +380,6 @@ INSERT INTO public.orders (id, customer_id, status, amount, note) VALUES (380, 514366, 's 381	178722	cancelled	2930.54	wmojdqnzik INSERT INTO public.orders (id, customer_id, status, amount, note) VALUES (382, 679322, 'c-383	668107	shipped	5325.74	zu INSERT INTO public.orders (id, customer_id, status, amount, note) VALUES (384, 442363, 'n

Guide: git integration. Use case: git says "Binary files differ".

MCP a diff an MCP client can query

rift-mcp lets an MCP client, such as an editor or a chat app, ask about the diff of two huge files without loading them. It has three tools, summary, hunks_at and next_difference, and every answer has a size limit.

$ rift-mcp
-> {"method":"tools/call","params":{"name":"summary","arguments":{"a":"a.sql","b":"b.sql"}}}
<- {
  "a_lines": 843633,
  "b_lines": 845034,
  "differs": {"a_lines":10936,"b_lines":12337,"segments":435},
  "equal": {"a_lines":832697,"b_lines":832697,"segments":435},
  "first_differences": [
    {"a":{"count":9,"start":378},"b":{"count":8,"start":378},"status":"differs"},
    {"a":{"count":21,"start":1006},"b":{"count":21,"start":1005},"status":"differs"},
    ...
  ],
  "mem_budget_bytes": 67108864,
  "more_differences": 415
}

Guide: MCP server.

VS Code the extension

Compare two huge files from VS Code and open any difference in VS Code's own diff editor.

Install from the Marketplace Install from Open VSX

Guide: the VS Code extension.

What a demo looks like summary, then the terminal view

1 GiB pair, full diff Apple M1

Apple M1, 16 GB RAM, internal SSD, macOS, warm page cache. Source: README.md at the v0.1.0 tag, section Benchmarks. All tables, with commands, are on the benchmarks page.

Apple M1 macOS, warm page cache

git diff has no 1 GiB row: it treats files over 512 MiB as binary and prints no diff. Peak memory is the resident set; Rift's heap peaked at 83.5 MiB on the 1 GiB pair and 94.1 MiB on the 10 GiB pair.

Intel Core i5-10600 Ubuntu 26.04.1 LTS, tmpfs

Median of three runs. GNU diff was killed at a 3 GiB cap on the 1 GiB pair. On the 1 GiB pair git diff printed no diff: it treats the files as binary. More rows on the benchmarks page.

What is guaranteed

• The output is a correct edit script: applied to the first file it gives exactly the second, byte for byte.
• A line counts as unchanged only after its bytes are compared. Hashes only propose matches.
• The heap stays at or under --mem. Files are read with pread into fixed buffers, never mmap.
• A region Rift could not refine within the budget is marked approximate in every output.

What is not

• The diff is not minimal.
• A moved block shows in the unified diff as a deletion and an insertion. --summary, --stats and the terminal view list moved blocks, and detection does not find every block.
• Windows is not supported.