Skip to content

Azure browser authentication helpers keep reading after callback disconnect #290

Open

Description

@saimanit

Test environment: Local Kali Linux x86_64, Python 3.14.8, connector checkout declaring version 2.2.0 at commit 248ffd5896c7a46bba115452dc10b8389d3ede54. The driver source was unchanged.

Affected code: run_server() in:

  • redshift_connector/plugin/browser_azure_credentials_provider.py
  • redshift_connector/plugin/browser_azure_oauth2_credentials_provider.py

Bug: When the callback peer disconnects, recv() returns b''. Neither helper checks for this condition, so it keeps reading the ended connection. Each read finishes immediately; the socket-operation timeout therefore does not stop the loop.

Debugging tools used on Kali Linux: I used socat, pidstat, and GNU timeout to debug and reproduce the disconnect-handling issue, together with Python's cProfile and pytest for profiling and follow-up checks. The reproduction executed the actual extracted methods using local Unix socket pairs and an accept() shim.

Tools and roles:

  • socat closed the peer endpoint without sending an authentication payload.
  • pidstat independently sampled the worker's CPU consumption.
  • Python's cProfile recorded the repeated native socket recv() calls.
  • GNU timeout imposed backup limits on worker/client process lifetimes.
  • pytest checked scripted callback behavior and the evidence parsers. These were not complete browser sign-in tests.

Python signal watchdogs interrupted the measured loops at the test budget.

Reproduction: With driver/ and reports/ beside each other, run from their parent directory:

timeout -k 2s 45s python3 reports/collect_callback_evidence.py

Observed: In approximately 2.3 seconds, Azure made 1,195,639 recv() calls and Azure OAuth2 made 785,928. Both used roughly one CPU core and required watchdog interruption. As a control, open, idle connections correctly timed out with negligible CPU use in both the original methods and the copies with the proposed fix applied in memory.

Impact if unfixed: The authentication helper can remain busy without returning a result or connection error. Applications waiting for it could have stalled sign-in attempts and sustained client CPU consumption.

Proposed fix: Check for EOF immediately after receiving data, inside the existing with conn: block, before decoding or parsing:

part = conn.recv(size)
if not part:
    raise InterfaceError("Connection closed before authentication completed")

In-memory testing of this change made both disconnected cases exit after one read in approximately 2 milliseconds, while idle timeouts still worked. The patch has not been applied to the driver checkout, and complete authentication regression testing remains necessary.

Evidence attachment: redshift-auth-evidence-minimal.zip contains nine files: the reproduction script, combined results, two CPU logs, two readable profiles, environment metadata, the proposed patch, and a README. The local username and machine-specific paths have been redacted; recorded measurements are preserved. Upload the ZIP with the report and use the attachment link generated by the reporting form.

How to review the files: Extract the ZIP and open the redshift-auth-evidence-minimal/ folder. Start with README.md for the setup and testing limits, then follow this structure:

redshift-auth-evidence-minimal.zip

redshift-auth-evidence-minimal/
├── README.md                         Reproduction instructions and testing limits
└── reports/
    ├── collect_callback_evidence.py   Script that reproduces the EOF/idle cases
    └── evidence/
        └── callback-eof-20261112T075315353916Z/
            ├── results.json          Outcomes and measurements for all 12 cases
            ├── environment.json      Python, platform, tools, and reviewed commit
            ├── azure-original-socat_eof.pidstat.txt     Azure CPU measurements
            ├── oauth2-original-socat_eof.pidstat.txt    OAuth2 CPU measurements
            ├── azure-original-socat_eof.profile.txt     Azure recv() call counts
            ├── oauth2-original-socat_eof.profile.txt    OAuth2 recv() call counts
            └── proposed-eof.patch    Suggested EOF check for both providers

Read results.json for the comparison, use the CPU logs and profiles to check the measured behavior, and review proposed-eof.patch for the suggested correction. The connector source is obtained separately at the reviewed commit and placed in driver/ beside reports/ before running the collector.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions