diff options
Diffstat (limited to 'docs/internal-server-errors.md')
| -rw-r--r-- | docs/internal-server-errors.md | 30 |
1 files changed, 30 insertions, 0 deletions
diff --git a/docs/internal-server-errors.md b/docs/internal-server-errors.md new file mode 100644 index 0000000..4f679b7 --- /dev/null +++ b/docs/internal-server-errors.md @@ -0,0 +1,30 @@ +# Internal Server Errors + +When `hi` encounters a problem that prevents a request from completing, it may report a `500 Internal Server Error` to clients, along with an error code. The actual error will be printed to standard error, with the error code. The following sections describe errors we've encountered, the likely operational consequences, and recommend approaches for addressing them. + +## database is locked + +The server attempted two write transactions at the same time, and encountered [sqlite's write locks](https://www.sqlite.org/rescode.html#busy). This is unfortunately unavoidable, but generally only occurs as a result of extremely bad luck, or very high load. + +This error will almost always resolve itself if clients re-try their requests; no further action is needed. + +This is a known issue. If you are encountering this consistently (or if you can trigger it on demand), let us know. We are aware of sqlite's features for mitigating this issue but have been unsuccessful in applying them; we're working on it, but patches _are_ welcome, if you have the opportunity. + +## stored canonical form […] does not match computed canonical form […] for name […] + +When `hi` applies the `migrations/20241019191531_canonical_names.sql` migration (from commit `3f9648eed48cd8b6cd35d0ae2ee5bbe25fa735ac`), this can leave existing names in a state where the stored canonical form is not the correct canonicalization of the stored display names of channels and logins. `hi` will abort requests when it encounters this situation, to avoid incorrect behaviours such as duplicate channels or duplicate logins. + +As channel and login names may be presented during client startup, this can render the service unusable until repaired. Treat this as an immediate outage if you see it. + +You can verify that login names are unique by running the following commands as the user the `hi` server runs as: + +* `sqlite3 .hi 'select display_name from login'` +* `sqlite3 .hi 'select display_name from channel_name'` + +Substitute `.hi` with the path to your `hi` database if it differs from the default. + +If the names are unique, you can repair the database: + +* Stop the `hi` server. +* Run `hi-recanonicalize`, as the same user the `hi` server runs as, with the same database options. +* Start the `hi` server. |