Merge queue planning: 40 merges in under 30 seconds

Before Merget lands a batch, it plans it: it works out which merges conflict with which, so independent groups can be prepared side by side. Planning can be spread over many workers. We planned 40 historical merges from scratch, with 1 worker and with 40, on two open-source repositories.

On Flask, 40 workers planned a 40-merge batch in 28.9 seconds against 241.9 with one worker: an 8.4× speedup. On Ripgrep it took 16.4 seconds against 213.6, about 13×.

MeasuredOct 29 – Nov 1, 2026
Workers1 and 40
Model spend$0, no model calls

Plan 40 merges in under 30 seconds. 8.4× faster on Flask with 40 workers; about 13× on Ripgrep.

8.4×faster on flask with 40 workers than with 18.38×: 250.7 and 233.0 s → 27.7 and 30.0 s
13.1×faster on ripgrep with 40 workers than with 113.07×: 213.6 s → 16.4 s
$0model spend: planning makes no model calls
Cold plan

Planning a 40-merge batch from scratch, with no warm index, without landing anything.

Workers

How many planning jobs run at once.

Speedup

The 1-worker time divided by the 40-worker time, each the median of 2 runs.

On ripgrep, planning 40 merges went from 213.6 seconds with 1 worker to 16.4 with 40. On flask, from 241.9 to 28.9 seconds.

How to read thisEach row joins the 1-worker time (hollow) to the 40-worker time (filled), median of 2 runs. The figure on the right is the speedup.

1 worker40 workers
Planning time, seconds
ripgrep741 candidate pairs, 669 conflict edges213.6 → 16.4 s13.07×
flask780 candidate pairs, 748 conflict edges241.9 → 28.9 s8.38×

Planning spreads well across workers on these two batches. With 40 workers, a 40-merge batch was planned in under 30 seconds instead of about four minutes.

With 40 workers, planning 40 merges took 16.4 seconds on ripgrep and 28.9 seconds on flask, against about three and a half and four minutes with one. Planning is then a small part of the time a batch takes to land.

Next step

Put a queue in front of your agents. Connect one repository and let Merget plan your open pull requests.

Sign up free
How we ran it

Measured October 29 to November 1, 2026: 40 historical merges each from ripgrep and flask, planned from scratch with 1 worker and with 40, median of 2 runs.