in_tail: add data reliability note

This commit adds a 'Data Reliability and Recovery' hint to the Tail input plugin documentation.

It clarifies the behavior of the database offset mechanism during unexpected shutdowns (e.g., system crash, power loss). Specifically, it explains that while Fluent Bit guarantees at-least-once delivery, there is a possibility of slight offset lag and minimal data duplication upon recovery. This ensures users understand that no data is lost even in these scenarios.

refs: fluent/fluent-bit#11269

Signed-off-by: jinyong.choi <[email protected]>

Author: jinyongchoi

pipeline/inputs/tail.md

@@ -89,6 +89,18 @@ If no database file is present, positioning behavior depends on the value of `re
 
 The database file essentially stores `inode=offset` so it should be unique per instance of the plugin, for example if you have two tail inputs then use two separate `db` files for each. That way each tail input can independently track its own state.
 
+{% hint style="info" %}
+
+**Data Reliability and Recovery**
+
+During normal operations and graceful shutdowns (SIGTERM), Fluent Bit synchronizes file offsets with the database to ensure no data is lost.
+
+In unexpected shutdowns (e.g., system crash, power loss, SIGKILL), Fluent Bit guarantees **at-least-once** delivery. The database offset may slightly lag behind the actual processed position if an **unexpected** shutdown occurs **after data is processed but before the new offset is committed** to the database.
+
+Upon restart, Fluent Bit resumes from the last committed checkpoint. This ensures **no data is lost**, though it may result in the re-ingestion of a minimal amount of previously processed records (duplication).
+
+{% endhint %}
+
 {% hint style="info" %}
 The `unicode.encoding` parameter is dependent on the `simdutf` library, which is itself dependent on C++ version 11 or later. In environments that use earlier versions of C++, the `unicode.encoding` parameter will fail.