Fixing Quiet Bugs in Open Source
Most of my 14 open-source fixes never raised an error. Three of them, from Apache Airflow and Argo CD, and why bugs like these are worth fixing.
· 6 min read · Open source, Debugging, Go, Python
The bugs that don't throw
Most of the bugs I fix in open source never raise an error. The program keeps running, the tests stay green, and the result is slightly wrong. A slice that should be empty returns rows. A pipeline takes the wrong branch, but only after midnight. A server keeps a little more memory every time someone closes a tab.
Since September 2025 I've written 14 fixes like this across 9 projects: Apache Airflow, Apache Superset, Argo CD, pg_lake, Ruby on Rails, Gin, Kubernetes, Grafana Loki and Firestone. This post walks through three of them.
Airflow: a slice that should have been empty
Airflow has an endpoint that slices a task's mapped outputs, and it is meant to behave exactly like a Python list. In Python, items[5:2] is an empty list. The endpoint translated the slice into SQLAlchemy's Query.slice(5, 2), which compiles to OFFSET 5 LIMIT -3.
What happened next depended on the database. SQLite treats a negative LIMIT as no limit at all, so it quietly returned every row from the fifth onward. PostgreSQL and MySQL rejected the query with a server error. Neither one returned the empty list Python would.
I found the code paths where the two bounds could cross, showed why the others were already safe, and made crossed slices return an empty result before any SQL is generated. I reproduced the bug with a script that only uses the standard library, then added regression tests for positive, negative and descending slices, so the behaviour is the same on all three databases.
Airflow again: the window after midnight
BranchDateTimeOperator picks a branch depending on whether the current time falls inside a window. With a window that crosses midnight, like 22:00 to 02:00, it worked before midnight. After midnight it compared the current time against the next night's window instead of the current one, so every run between midnight and 02:00 took the wrong branch, with nothing in the logs.
The fix checks which side of midnight the current time is on and anchors the window to the right day. The test covers times before midnight, after midnight and exactly on the boundary. The after-midnight cases fail without the fix.
Argo CD: one word, two release branches
In Argo CD's pod log streaming, when sending a log line to a client failed, the goroutine wrote the error to a one-slot channel and then hit break. That led to a second send on the same channel. If the client had already disconnected, nothing was left to read it, so the goroutine blocked forever, holding its stream. argocd-server leaked one goroutine every time someone closed a log view while logs were still streaming.
An earlier fix had made the channel buffered so the goroutine could always send and exit, but that only works if it sends once. The fix was to change break to return, matching the error path just above it. The test simulates a client that disconnects mid-stream and checks that no streaming goroutine is left running. It fails without the change and passes on repeated runs. The fix was merged and cherry-picked into the Argo CD 3.5 and 3.6 release branches.
Why they matter
Quiet bugs are worth the time because nobody is looking for them. A crash gets reported within a day. A filter that drops the value 0, or a log path that still contains a phone number, can sit in a release for months.
Each fix has its own write-up on the open-source pages of this site, with the problem, the fix and how I tested it.