The TRS spec mentions server rollbacks in several places, but never defines what these are. A definition should be added. There is actually no concrete representation for a rollback event. Instead, a server rollback is inferred when the client detects certain conditions. The spec [1] has the following text:
"In the (hopefully rare) situation that the Client fails to find its sync point event, one of two things is likely to have happened on the Server: either the Server has truncated its Change Log, or the Server has been rolled back to an earlier state. If the Client had been retaining a local record of previously processed events, the Client may be able to detect a Server rollback if it notices the successor event of some previously processed event has been removed or changed to one with a different identifier than before." My dev team is working with a client implementation of the TRS spec (LQE) that interprets certain contains in the TRS feed as indicating a rollback event, and then re-indexes the entire data source. This behavior is undesirable since indexing a large data source can take days, during which time users can't get accurate query results. I recommend that we expand the guidance for how TRS clients should respond to an inferred rollback event. There should be other less disruptive courses of action. In some cases the rollback event is caused by other factors. We have observed that the spec is difficult to implement unless the server maintains certain information, e.g. a record of each change. In our experience, we have never actually rolled back our server, but due to race conditions we occasionally produce a change log that appears to contain a rollback event. The alternate responses to a rollback include: 1. ignore - the client continues to process the change log and makes a sensible guess about where to cut off, e.g. by remembering some information from the previous change log 2. halt - the client stops processing and waits for an administration to explicitly select the next action which could be ignore or re-index The client should be configured with a suitable policy, e.g. ignore, halt, or re-index, and have an admin interface so that a human administrator can take the best course of action. In any case, a unilateral automatic decision to re-index is problematic. Another way to deal with rollback events is to add a new type of event to the change log, i.e. a trs:Rollback event. Only when this event is received should a client re-index. Minor point: the text of the specification should not use both the terms "cutoff event" and "synch point". Let's pick one and use it throughout. Regards, ___________________________________________________________________________ Arthur Ryman DE, Chief Architect, Reporting & Portfolio and Strategy Management IBM Software, Rational Toronto Lab | +1-905-413-3077 (office) | +1-416-939-5063 (mobile)
