---
bc-version: [all]
domain: error-handling
keywords: [logging, rollback, session, transaction, isolated-session, telemetry]
technologies: [al]
countries: [w1]
application-area: [all]
---

# Log writes that must capture failures must survive transaction rollback

## Description

Inserting a log record inside the same transaction as the operation it logs looks correct until the operation errors: the transaction rolls back and takes the log entry with it. The result is a log that faithfully records every success and silently loses exactly the failures it exists to capture. This is a common blind spot in error/duration logging around web-service calls, background jobs, and other operations expected to fail sometimes.

## Best Practice

Write any log whose purpose includes capturing failures from a transaction that is independent of the operation being logged: start an isolated session (`Session.StartSession` on a `TableNo`-scoped codeunit that only inserts the log record and commits) so the entry persists regardless of what happens to the caller's transaction. `StartSession`'s only channel for getting data into that new session is its optional `Record` parameter, delivered to the target codeunit's `OnRun` trigger — a separate session gets a fresh instantiation of the codeunit, so calling a setter procedure on a local object variable before starting the session does not populate anything in the new session's instance. Capture duration and other telemetry values in the caller, place them into the `Record` passed to `StartSession`, and do the insert-and-commit entirely inside that session's own `OnRun`. Logs that only record successful, committed work can safely stay in the main transaction; this pattern targets error and diagnostic logs specifically.

See sample: [`log-writes-must-survive-rollback.good.al`](log-writes-must-survive-rollback.good.al).

## Anti Pattern

Inserting the error-log record in the same transaction as the risky operation, so a rollback deletes the very entry meant to explain the failure. Adding a stray `Commit` before the risky call is not a fix either — it breaks the caller's atomicity and can violate posting-routine rules. Swallowing the error to keep the log alive (running a codeunit without checking or re-raising its result) is equally wrong: the log must observe the failure, not suppress it.

See sample: [`log-writes-must-survive-rollback.bad.al`](log-writes-must-survive-rollback.bad.al).