diff --git a/.github/runner-scale-sets/query-regression/README.md b/.github/runner-scale-sets/query-regression/README.md
index 3e0159af60..f7a9bdc173 100644
--- a/.github/runner-scale-sets/query-regression/README.md
+++ b/.github/runner-scale-sets/query-regression/README.md
@@ -51,10 +51,10 @@ Configuration lives in repository variables/secrets:
 | vars | `QUERY_REGRESSION_ECS_IMAGE_ID` | Custom image built by `ecs-image/build-ecs-image.py`. |
 | vars | `QUERY_REGRESSION_COMMENT_ALLOWLIST` | Comma/whitespace-separated GitHub logins allowed to comment `/query-regression` on a PR. Each login must also have repository `admin` permission. Empty denies all comment commands. |
 
-The system disk is 50 GiB, which covers the image, a 16 GiB swapfile, the
-checkout, and cold build caches (target dir, cargo registry, sccache). ENOSPC
-stops the runner itself from writing logs, which GitHub reports as `The
-operation was canceled` with no telemetry, indistinguishable from a
+The system disk is 80 GiB, providing capacity headroom for the image, a 16 GiB
+swapfile, the checkout, and cold build caches (target dir, cargo registry,
+sccache). ENOSPC stops the runner itself from writing logs, which GitHub reports
+as `The operation was canceled` with no telemetry, indistinguishable from a
 platform-side cancellation.
 cloud-init masks `systemd-oomd`, disables `unattended-upgrades` /
 `apt-daily-upgrade`, creates `/swapfile`, and sets `OOMPolicy=continue`
@@ -180,7 +180,7 @@ reintroduce those actions unless the image contract changes.
 Each run provisions its own ECS instance, so overlapping dispatches proceed
 in parallel. There is no workflow `concurrency` group. All Cargo state
 (`CARGO_HOME` including registry/git, `CARGO_TARGET_DIR`, sccache, cache
-metadata) lives on the 50 GiB system disk and is discarded with the VM.
+metadata) lives on the 80 GiB system disk and is discarded with the VM.
 `RUSTUP_HOME=/opt/rustup` and `/opt/cargo/bin` are image-owned. The runner
 sets `RUSTC_WRAPPER=/usr/local/bin/sccache`,
 `SCCACHE_DIR=/home/runner/.cache/sccache`, `SCCACHE_CACHE_SIZE=10G`, and
@@ -189,7 +189,7 @@ at 10G. Base and candidate builds share the target on that disk; Cargo
 fingerprints invalidate source and dependency changes.
 
 The workflow reports `du` / `df` in telemetry. It does not try to reclaim
-space across runs: a cold 50 GiB disk that fills up fails the build.
+space across runs: a cold 80 GiB disk that fills up fails the build.
 
 ## Future optional phases
 
diff --git a/.github/scripts/aliyun-ecs-runner-provision.py b/.github/scripts/aliyun-ecs-runner-provision.py
index 307ee336eb..b5dee7f51f 100644
--- a/.github/scripts/aliyun-ecs-runner-provision.py
+++ b/.github/scripts/aliyun-ecs-runner-provision.py
@@ -81,8 +81,8 @@ POLL_INTERVAL_SECONDS = 5
 SWAP_FILE = "/swapfile"
 SWAP_SIZE_GIB = 16
 # Image (~12G), 16G swap, checkout, base+candidate data homes, and cold
-# build caches. Deleted with the instance.
-SYSTEM_DISK_GIB = 50
+# build caches. 80 GiB provides capacity headroom. Deleted with the instance.
+SYSTEM_DISK_GIB = 80
 
 
 @dataclass(frozen=True)