Presigned URL support for S3 log files by rasmusfaber · Pull Request #3324 · UKGovernmentBEIS/inspect_ai

rasmusfaber · 2026-02-25T12:26:44Z

This PR contains:

What is the current behavior? (You can also link to an open issue here)

Large S3 hosted .eval-files can be very slow to display, and there is no progress bar.

What is the new behavior?

The server and frontend now supports use of presigned URLs on an opt-in basis. (Using the presigned URLs will require correct CORS setup on the S3 bucket, so it should not be indiscriminately applied).
Also, for large files, use parallel chunk fetching (up to 6 concurrent requests of 8MB per chunk).
Show progress bar for large samples.
Skip the decode()/encode() round-trip when parsing data by sending the raw bytes directly to the Web Worker.

Does this PR introduce a breaking change? (What changes might users need to make in their application due to this PR?)

The /log-size/ API endpoint is replaced by /log-info/ with a different response shape (JSON object instead of bare integer). The VSCode extension JSON-RPC method is renamed from eval_log_size to eval_log_info, but the frontend includes a backward-compatible fallback.

Other information:

src/inspect_ai/_view/server.py

dragonstyle · 2026-03-06T14:36:00Z

src/inspect_ai/_view/www/src/app/navbar/ApplicationNavbar.tsx

      return !!loading || sampleStatus === "loading";
    } else if (showActivity === "log") {
-      return !!loading;
+      return !!loading || sampleStatus === "loading";


This seems odd (specifically showing 'log' activity when samples are loading) - was this truly needed?

The hasActivity flag gates the ActivityBar rendering entirely. If it's false, neither the determinate progress bar nor the indeterminate spinner renders, even if sampleProgress has a valid value.

When a user is in the log view (showActivity="log") and clicks a sample to load, loading is already false (the log finished loading), so only sampleStatus === "loading" keeps the navbar progress bar visible during the download. Removing it would hide the progress bar for sample downloads in the log view.

There's also an edge case: when a log has exactly 1 sample, it's displayed inline via InlineSampleDisplay (which has its own ActivityBar), but the navbar indicator would still disappear.

But perhaps there is something that needs renaming or a comment here?

The original model would've been that the 'header' region of the sample would show immediately, since it is renderable using only a sample summary which should be present before the sample is loaded. This view as a sample responsive activity bar between the header region and the sample contents - the intent is that this conveys the loading of the sample.

This seems to be working as intended for inline samples, but non-inline sample display appears to be not behaving this way... I'm looking into this separately....

I've fixed this issue here:
#3444

The behavior changed when I switched us to using the SampleDetailView (and it was incorrect). This correctly shows the sample header immediately and shows progress beneath that header when samples load.

Thanks, that is cleaner.

I have removed the changes to ApplicationNavBar.tsx.

src/inspect_ai/_view/www/src/app/types.ts

src/inspect_ai/_view/www/src/client/remote/remoteLogFile.ts

src/inspect_ai/_view/common.py

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Replace get_log_size with get_log_info across view-server, static-http, and vscode API clients. The new LogInfo type includes an optional direct_url field for S3 presigned URLs. Co-Authored-By: Claude Opus 4.6 <[email protected]>

When log-info returns a direct_url (S3 presigned URL), byte-range requests go directly to S3. Falls back to server proxy on failure. Co-Authored-By: Claude Opus 4.6 <[email protected]>

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Replaced by log-info which returns size + optional direct_url. Co-Authored-By: Claude Opus 4.6 <[email protected]>

Co-Authored-By: Claude Opus 4.6 <[email protected]>

S3 buckets with SSE-KMS require SigV4. Add config_kwargs with signature_version s3v4 to the default S3 filesystem options. Co-Authored-By: Claude Opus 4.6 <[email protected]>

Eliminate the 30-byte header pre-fetch in readFile by estimating the total fetch size from the central directory's filenameLength + 256 bytes of extra field padding, then fetching once. Re-fetches only if the actual extra field exceeds the estimate (rare). Also cache the opened ZIP in get_log_details so subsequent get_log_sample calls reuse it. Co-Authored-By: Claude Opus 4.6 <[email protected]>

Add asyncJsonParseBytes() that transfers raw Uint8Array directly to the Web Worker pool, eliminating ~800ms of main-thread string allocation for large eval log files (~20MB). Co-Authored-By: Claude Opus 4.6 <[email protected]>

Thread onProgress callback from parallel chunk fetcher through the remote ZIP/log file readers to the UI, showing download progress for samples over 20MB instead of only an indeterminate animation. Co-Authored-By: Claude Opus 4.6 <[email protected]>

- Wrap presigned URL generation in try/except so failures degrade gracefully instead of crashing the /log-info/ endpoint - Move ProgressCallback type to client/api/types.ts to fix inverted dependency (api layer was importing from remote/remoteZipFile) - Move fetchRange to utils/http.ts canonical location, re-export from remoteZipFile for backward compat - Simplify fetchBytes wrapper in remoteLogFile and add console.warn on presigned URL fallback for CORS debugging - Deduplicate JSON/JSON5 parse fallback in asyncJsonParseBytes - Fix progress double-counting on ZIP re-fetch path - Guard against division by zero in progress display - Add truncated ZIP header length check - Add VSCode backward compat: fall back to eval_log_size if extension doesn't support eval_log_info yet Co-Authored-By: Claude Opus 4.6 <[email protected]>

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Lower parallel chunk threshold from 20MB to 8MB and increase max concurrent chunks from 6 to 10. Benchmarking across 20-200MB files showed 8MB chunks with 10 concurrency achieves ~4x speedup over single requests for 200MB files (cross-region). Co-Authored-By: Claude Opus 4.6 <[email protected]>

- Restore /api/log-size endpoint for VS Code extension compatibility - Rename DownloadProgress to Progress with complete/total fields - Pass content length directly to openRemoteZipFile instead of wrapper fn - Add clarifying comment for fsspec async _url convention Co-Authored-By: Claude Opus 4.6 <[email protected]>

Sample download progress is handled by InlineSampleDisplay's own ActivityBar, so the navbar bar should only show log-level loading. Co-Authored-By: Claude Opus 4.6 <[email protected]>

rasmusfaber changed the title ~~feat: presigned URL support for S3 log files~~ Presigned URL support for S3 log files Feb 25, 2026

rasmusfaber force-pushed the faber/presigned-url branch from 00fb032 to 1d8c4d0 Compare February 25, 2026 12:35

rasmusfaber marked this pull request as ready for review February 25, 2026 12:56

rasmusfaber mentioned this pull request Feb 25, 2026

Add server-side /log-details and /log-sample endpoints #3213

Closed

4 tasks

rasmusfaber force-pushed the faber/presigned-url branch from ceb958a to 19e2c66 Compare February 26, 2026 20:04

rasmusfaber marked this pull request as draft February 26, 2026 20:07

rasmusfaber force-pushed the faber/presigned-url branch from 8599bf8 to bda4b12 Compare February 26, 2026 20:13

rasmusfaber marked this pull request as ready for review February 26, 2026 20:14

rasmusfaber force-pushed the faber/presigned-url branch from c93d201 to 97a388c Compare February 27, 2026 09:35

dragonstyle self-requested a review March 4, 2026 19:10

rasmusfaber force-pushed the faber/presigned-url branch from 3734d92 to 8039504 Compare March 5, 2026 15:42

dragonstyle requested changes Mar 6, 2026

View reviewed changes

rasmusfaber force-pushed the faber/presigned-url branch from 345fc08 to f717fe3 Compare March 9, 2026 15:50

rasmusfaber requested a review from dragonstyle March 9, 2026 15:50

rasmusfaber and others added 16 commits March 10, 2026 12:59

feat: add log-info endpoint with optional direct_url support

1666c4a

Co-Authored-By: Claude Opus 4.6 <[email protected]>

feat: add log-info endpoint to aiohttp server

cea9943

Co-Authored-By: Claude Opus 4.6 <[email protected]>

feat: use direct_url for byte-range fetches when available

68427c3

When log-info returns a direct_url (S3 presigned URL), byte-range requests go directly to S3. Falls back to server proxy on failure. Co-Authored-By: Claude Opus 4.6 <[email protected]>

test: add test for generate_direct_urls with non-S3 files

22c1524

Co-Authored-By: Claude Opus 4.6 <[email protected]>

refactor: remove deprecated log-size endpoint

22fdab1

Replaced by log-info which returns size + optional direct_url. Co-Authored-By: Claude Opus 4.6 <[email protected]>

build: rebuild frontend with log-info support

4992856

Co-Authored-By: Claude Opus 4.6 <[email protected]>

fix: use S3 Signature V4 for presigned URLs

37aaa5c

S3 buckets with SSE-KMS require SigV4. Add config_kwargs with signature_version s3v4 to the default S3 filesystem options. Co-Authored-By: Claude Opus 4.6 <[email protected]>

docs: add presigned URL support to changelog

3920682

Co-Authored-By: Claude Opus 4.6 <[email protected]>

fmt

3f48027

rasmusfaber force-pushed the faber/presigned-url branch from f717fe3 to 2369c8e Compare March 10, 2026 12:50

fix: revert navbar ActivityBar to log-level only

ddce505

Sample download progress is handled by InlineSampleDisplay's own ActivityBar, so the navbar bar should only show log-level loading. Co-Authored-By: Claude Opus 4.6 <[email protected]>

rasmusfaber force-pushed the faber/presigned-url branch from 4132bb4 to ddce505 Compare March 10, 2026 12:57

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Presigned URL support for S3 log files#3324

Presigned URL support for S3 log files#3324
rasmusfaber wants to merge 17 commits intoUKGovernmentBEIS:mainfrom
METR:faber/presigned-url

rasmusfaber commented Feb 25, 2026 •

edited

Loading

Uh oh!

Uh oh!

dragonstyle Mar 6, 2026

Uh oh!

rasmusfaber Mar 9, 2026

Uh oh!

dragonstyle Mar 9, 2026

Uh oh!

dragonstyle Mar 9, 2026

Uh oh!

rasmusfaber Mar 10, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

rasmusfaber commented Feb 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

This PR contains:

What is the current behavior? (You can also link to an open issue here)

What is the new behavior?

Does this PR introduce a breaking change? (What changes might users need to make in their application due to this PR?)

Other information:

Uh oh!

Uh oh!

dragonstyle Mar 6, 2026

Choose a reason for hiding this comment

Uh oh!

rasmusfaber Mar 9, 2026

Choose a reason for hiding this comment

Uh oh!

dragonstyle Mar 9, 2026

Choose a reason for hiding this comment

Uh oh!

dragonstyle Mar 9, 2026

Choose a reason for hiding this comment

Uh oh!

rasmusfaber Mar 10, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

rasmusfaber commented Feb 25, 2026 •

edited

Loading