Compact scrollback in Alacritty

The ingest decision left compact history as a technical go whose adoption depends on the engine. This report records the attempt to take the design upstream: an opt-in port to Alacritty master, measured in Alacritty itself against the unmodified revision.

Upstream closed the proposal, alacritty/alacritty#9049, and no pull request was opened. The port is archived here as final evidence. It is not used by Rozi.

Candidate and baseline

• Baseline: Alacritty d692748d, release build.
• Candidate: the same revision with the port, archived as tools/experiments/alacritty-port. Its last commit is e8e5aee6 on a local branch.
• dense is the candidate with compaction off, which is what every existing Alacritty user would run. compact sets ALACRITTY_COMPACT_SCROLLBACK=1.

The port keeps negative-line Grid indexing for dense rows and panics only on a compact row, which must be read through Grid::read_row. The candidate passes 148 alacritty_terminal unit tests, 45 reference tests, and a differential test that compares mixed output and resizes with compaction off, on, and toggled against alacritty_terminal 0.26.0.

AMD Ryzen 7 5700X3D, Linux 7.2.3, Wayland, Rust 1.98.1. Each run waited for load average below 2 and no running rustc.

What was measured

• vtebench in a fullscreen 284×68 window with scrolling.history = 10000 and --max-secs 5, plus short_lines, 2,000 log lines of about 40 columns. Upstream, compact, and dense alternate. The table gives the median of per-round medians. vtebench reports whole milliseconds, so one tick on a 9 ms benchmark reads as 11%.
• Process PSS after 12,000 lines of short log output or full-width random text, sampled while idle.
• Criterion benchmarks of the engine alone, at 253×64 with 10,000 lines of history, interleaved over 5 rounds against a baseline built from the same benchmark file.

Alacritty

Three rounds, with every fix except the cursor-cell change:

vtebench, ms per sample	upstream	dense	compact
short_lines	18	16	14
scrolling_fullscreen	9	9	10
dense_cells	93	99	99
cursor_motion	32	33	33
scrolling	144	146	146
scroll regions, four benchmarks	143–147	144–148	144–145
light_cells, medium_cells, sync_medium_cells, unicode	8–12	8–12	8–12, equal to upstream

Five rounds of dense_cells alone after the cursor-cell change:

dense_cells	upstream	dense	compact
median of round medians, ms	92	95	94
round medians, ms	92, 92, 96, 93, 92	95, 95, 94, 95, 94	94, 94, 94, 94, 95
mean of all samples, ms	93.3	94.7 (+1.6%)	94.3 (+1.1%)

PSS after filling history, MiB	upstream	dense	compact
short lines	116.9	117.1	59.5
full-width lines	121.7	122.0	121.8

With short lines in history, compact storage halves Alacritty's PSS, and anonymous memory falls from 89.3 to 32.0 MiB. Full-width history is the same size in every build.

Compact mode is also the fastest build on short_lines. The engine benchmarks below charge it about 30% more CPU per short line, and the application does not show that cost. Writing and touching less memory per scrolled line is a likely reason, but this run does not isolate it.

Engine

Criterion case	upstream	dense	compact
ingest_100k/short	21.7 ms	+3.6%	+29.5%
ingest_100k/full	182.6 ms	−1.7%	−0.3%
ingest_100k/uniform	182.2 ms	−1.1%	−0.6%
alt_screen/dense_cells	55.2 ms	+0.2%	+0.7%
display_iter/live	28.2 µs	−19.0%	−17.1%
display_iter/top	28.2 µs	−17.2%	−26.6%
iterate_history	4.63 ms	−15.3%	−29.3%

The first four rows are measured at e8e5aee6. The read rows are measured at 6eb5f322; the later commits do not change read paths and were not measured on them.

Regressions found and fixed

Each was found by Alacritty or engine benchmarks, reproduced in the engine, and fixed before the numbers above:

• display_iter was 55–98% slower because the iterator resolved a history row per cell. It now resolves a row once per line.
• Dense ingest with full history went through the row pool. A full dense history now swaps the oldest row with the scrolled viewport row.
• The negative-line check on Grid indexing stopped the per-character write path from inlining. History indexing moved to cold helpers.
• Rows of one repeated glyph across the full width compacted to a single cell only after a full-row comparison: +21% in the engine and +33% in vtebench scrolling_fullscreen. Such rows now stay dense. The engine case is at −0.6%, and Alacritty went from 12 to 10 ms against upstream's 9 ms.
• Writing to the cursor cell went through the history check on every character: dense_cells was +3.6% dense and +4.4% compact in the engine, and 93 → 99 ms in Alacritty. Indexing the viewport directly brings the engine to +0.2% and +0.7%, and Alacritty to about +1–2%.

Shrinking a history row from 64 to 40 bytes did not move the remaining +3.6% dense short-line engine cost. A 32-byte row, which needs bit-packing and a 64-bit-only layout, was not tried, since nothing points to row size as the cause.

Result

Compact scrollback is a credible production design. Against unmodified Alacritty it halves process memory with short-line history and is faster on short-line output. The other vtebench workloads stay within 3% of upstream, apart from one tick on the 9 ms scrolling_fullscreen. With compaction off, the port matches upstream except for the +3.6% short-line engine case, which Alacritty does not show, and about 1.5 ms (+1.6%) in dense_cells.

Upstream will not take it, so Rozi can use it only through an alacritty_terminal fork maintained alongside tui-lipan. Whether the memory saving justifies that fork is a separate product decision.

Not measured: resize and reflow in Alacritty, and the port inside Rozi.