Registry logs show Compose v5.5.1 ignoring pull_policy windows on explicit pulls

Under `up`, the saving comes from a time check against the local image. With `pull_policy: every_3s`, an immediate second `up` skipped the registry, and a third `up` after a six-second sleep checked again and re-pulled ^[8]. The window expires. The author ran that test to rule out a permanent "already pulled" flag ^[8]. Compose has accepted `daily`, `weekly` and `every_<duration>` since v2.34.0 ^[2]. The v5.1.1 plugin bundled with Engine 29.3.1 is well past that release, so it already applies the windows during `up` ^[3].

The explicit pull is where the test and the release notes disagree. The v5.5.0 changelog, released 17 August 2026, says `compose pull` now respects the daily, weekly and every_N refresh windows set in pull_policy ^[3]. The developer who checked it, writing on dev.to, downloaded the v5.1.1 and v5.5.1 binaries from GitHub ^[5]. Both ran against a local `registry:2` container holding one small image, with its access log tailed to count manifest checks ^[5]. It is a clean harness. The count comes from the registry's own log, and the cold start used `docker rmi -f` so no copy of the image stayed on the daemon ^[13].

The run covered the case the changelog names: plain window values, an explicit `pull`, nothing else in play ^[16]. Swapping `daily` in for `every_3s` gave the same result ^[10]. The author saw no difference in behaviour between the pre-fix and post-fix binaries ^[16]. This is one developer's harness against one local registry, and it is enough to establish that the changelog's named case behaved the same on both binaries there. "I can't tell from the outside what the fix actually touched," the author wrote ^[15]. On CI the author was plain: "If you were hoping the fix meant your CI pipeline's docker compose pull step would start skipping unnecessary registry calls, it doesn't, at least not in this scenario." ^[14]

The 80 percent cut under `up` ^[1] comes from the author's workload: five back-to-back cycles inside one window on a machine that kept the image. It transfers only to a pipeline that calls `up` more than once per window on a runner that keeps its image store between calls ^[5]. A runner that starts empty repeats the cold first cycle, and in the test that cycle checked the registry ^[1]. Locally the saving did not show on the clock. Both five-cycle runs took about 53 seconds, mostly container and network setup ^[6]. Against Docker Hub, four avoided round trips at the measured 0.40 to 1.01 seconds each come to about 1.6 to 4.1 seconds per five cycles ^[2]. The author wrote that "on a slow or rate-limited registry it adds up over a day of repeated up calls" ^[17].

The window also costs freshness. "A refresh window that skips checks also skips seeing real changes," the author wrote ^[12]. Because the skip is time-based, a new image pushed to the same tag inside the window goes unchecked by `up` until the window runs out ^[4].