Project

General

Profile

Feature #48

Scanning for corrupt indices at gdplogd startup

Added by Nitesh Mor about 4 years ago. Updated over 1 year ago.

Status:
Closed
Priority:
Low
Assignee:
Category:
gdplogd
Start date:
08/23/2016
Due date:
% Done:

0%


Description

Even with appropriate flushing of appends to data file and to index files (recno=>offset and timstamp=>recno), there is a chance of application/system crash at a specific moment which will end up in out-of-sync condition for the index files.

The root cause of this is that the update to data file and the index files is not an atomic operation. It'd be nice to have the ACID semantics for such updates, but one could argue that the index files could always be rebuilt from the data file, and hence the strong ACID semantics aren't needed for the entire operation. However, this leads to the necessity for detecting such corruption and repairing it.

There are various ways to do it, here is just one possible way: The corruption is limited to only one record (assuming the data is flushed with each append individually). At startup time, gdplogd has to scan the local disk for the available records anyways. An extra step would be to verify the integrity of each index file by just looking at the most recent entry, and if needed, do a repair. Another optimization could be to keep a soft-state on persistent storage to mark non-corrupt logs on a clean shutdown.


Related issues

Related to GDP - Feature #46: Need way of rebuilding log indices ("fsck for logs") Closed 08/22/2016

History

#1 Updated by Nitesh Mor about 4 years ago

  • Related to Feature #46: Need way of rebuilding log indices ("fsck for logs") added

#2 Updated by Eric Allman over 1 year ago

  • Status changed from New to Closed

Not relevant for GDPv2, which uses SQLite for log storage.

Also available in: Atom PDF