[LU-11471] IO Errors during failover with very few number of OSTs

• Status: Closed
• Project: Lustre
• Type: Bug
• Priority: Minor
• Reporter: Rajeshwaran Ganesan
• Assignee: Mikhail Pershin
• Resolution: Duplicate

Environment

Lustre 2.5.X

• Severity: 3

Description

one of our customer is noticed IO errors during failover with fewer OST. we would like to add a note into the Best practice section or OST planning section
========================
During Fail over situation with few OSTs possibly cause I/O errors on the client side.

Issue Links

• is related to: LU-10995 DoM2: allow MDT-only filesystems

Comments

[Rajeshwaran Ganesan]

we would like add, If the MDS has no currently active OSTs, create requests fail with an I/O error. And we will see the IO errors at the clients.

[Joseph Gmitter (Inactive)]

Hi Rajeshwaran,

Are you familiar with how to push such an update to the manual?

For details on how to submit changes to the manual, please see:

• https://wiki.hpdd.intel.com/display/PUB/Making+changes+to+the+Lustre+Manual
• https://wiki.hpdd.intel.com/display/PUB/Code+Documentation+Guidelines

[Andreas Dilger]

I wonder whether this should rather be considered a bug in the code, and the MDS should block file creations if all of the OSTs become unavailable after startup?

[Rajeshwaran Ganesan]

please close this case

[Andreas Dilger]

Reopening this issue. With the advent of Data-on-MDT we will at some point want to allow filesystems with only MDTs to be created. At that point, this check has to be removed.

As a starting point, we could add a tunable that allows this behavior to be selected by the admin - return an error if no OSTs are available, or cause the client to block and wait for an OST to become available. I think in the case where an OST was previously available, but they are temporarily offline due to failover, the client should block. If the file being created has a DoM component at the start, then it should not block.

[Mikhail Pershin]

Main ticket for remaining work is LU-10995