github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-06 16:11:34 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #15651] create: blobs can still be GC-pruned during long-running create operations after the 1h grace window #56500

New Issue
Open
opened 2026-04-29 10:55:51 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-04-29 10:55:51 -05:00
Owner
Copy Link

Originally created by @shaun0927 on GitHub (Apr 17, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/15651

What is the issue?

PR #15628 reduces the create/GC race by skipping pruning for blobs newer than 1 hour, but the failure mode still exists for create/import flows that run longer than that window.

Right now the protection is still wall-clock based:

  • server/images.go sets layerPruneGracePeriod = time.Hour
  • PruneLayers() skips blobs only while time.Since(info.ModTime()) < layerPruneGracePeriod
  • manifest/layer.go refreshes mtimes when layers are created or reused, but there is no durable in-progress signal while a long-running create is still active

That means a create flow can still lose pending blobs if:

  1. blobs have already been written,
  2. the manifest has not been written yet, and
  3. another Ollama process runs GC after those blob mtimes age past the 1 hour grace window.

I validated the mechanism locally against current main (57653b8e42d69ec35f68a59857bad4d0f07994a3) by:

  1. creating a blob with manifest.NewLayer(...)
  2. aging the blob mtime beyond the grace window
  3. calling PruneLayers()
  4. confirming the blob is deleted
  5. confirming a later manifest.NewLayerFromLayer(...) fails because the blob is gone

Expected behavior: blobs that are still part of an active create/import should not become GC-eligible solely because more than 1 hour elapsed before the manifest was written.

Impact: long-running create/import/quantize flows can still fail late and lose work.

Relevant log output

=== RUN   TestValidation_LongCreateBlobCanBePrunedAfterGraceWindow
time=2026-04-18T02:08:44.323+09:00 level=INFO msg="total blobs: 1"
time=2026-04-18T02:08:44.323+09:00 level=INFO msg="total unused blobs removed: 1"
--- PASS: TestValidation_LongCreateBlobCanBePrunedAfterGraceWindow (0.00s)

OS

macOS

GPU

Apple

CPU

Apple

Ollama version

main @ 57653b8e42d69ec35f68a59857bad4d0f07994a3 (v0.21.0 tag points at the same commit)

Originally created by @shaun0927 on GitHub (Apr 17, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/15651 ### What is the issue? PR #15628 reduces the create/GC race by skipping pruning for blobs newer than 1 hour, but the failure mode still exists for create/import flows that run longer than that window. Right now the protection is still wall-clock based: - `server/images.go` sets `layerPruneGracePeriod = time.Hour` - `PruneLayers()` skips blobs only while `time.Since(info.ModTime()) < layerPruneGracePeriod` - `manifest/layer.go` refreshes mtimes when layers are created or reused, but there is no durable in-progress signal while a long-running create is still active That means a create flow can still lose pending blobs if: 1. blobs have already been written, 2. the manifest has not been written yet, and 3. another Ollama process runs GC after those blob mtimes age past the 1 hour grace window. I validated the mechanism locally against current `main` (`57653b8e42d69ec35f68a59857bad4d0f07994a3`) by: 1. creating a blob with `manifest.NewLayer(...)` 2. aging the blob mtime beyond the grace window 3. calling `PruneLayers()` 4. confirming the blob is deleted 5. confirming a later `manifest.NewLayerFromLayer(...)` fails because the blob is gone Expected behavior: blobs that are still part of an active create/import should not become GC-eligible solely because more than 1 hour elapsed before the manifest was written. Impact: long-running create/import/quantize flows can still fail late and lose work. ### Relevant log output ```shell === RUN TestValidation_LongCreateBlobCanBePrunedAfterGraceWindow time=2026-04-18T02:08:44.323+09:00 level=INFO msg="total blobs: 1" time=2026-04-18T02:08:44.323+09:00 level=INFO msg="total unused blobs removed: 1" --- PASS: TestValidation_LongCreateBlobCanBePrunedAfterGraceWindow (0.00s) ``` ### OS macOS ### GPU Apple ### CPU Apple ### Ollama version main @ `57653b8e42d69ec35f68a59857bad4d0f07994a3` (`v0.21.0` tag points at the same commit)
GiteaMirror commented 2026-04-29 10:55:53 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Apr 17, 2026):

A workaround is to set OLLAMA_NOPRUNE=1 in all but one of the servers accessing the model directory.

<!-- gh-comment-id:4270309632 --> @rick-github commented on GitHub (Apr 17, 2026): A workaround is to set `OLLAMA_NOPRUNE=1` in all but one of the servers accessing the model directory.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#56500
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?