Continuous Replication jobs creates full backups every time since 2025-09-06 (xo from source)

olivierlambert

You can try to pinpoint the exact commit to be 100% sure, with git bissect

peo

@olivierlambert said in Continuous Replication jobs creates full backups every time since 2025-09-06 (xo from source):

git bissect

That might be quite a bunch of commits in between the working one and when I discovered the problems (I actually let it do it's thing a couple of days, quickly filling up the backup drive)
I keep only the current and the previous build of XO, so I only know it's somewhere between Aug 16 (from the 'lastlog' to find out when I was logged in) and late Sep 5.

olivierlambert

Bissect is rather fast, it… bissects between 2 commits. If you can pinpoint the exact commit generating that, it means we could probably solve this a LOT faster

For ref: https://en.wikipedia.org/wiki/Bisection_(software_engineering)

peo

@olivierlambert Has the problem been confirmed ? I can try to pinpoint where the always-full replications were introduced, but more time efficient to do it on a machine that is fast do replicate, not the one that now takes 2.5 hours (a few minutes for the diff before it borked starting at the backup made 6 Sept)

olivierlambert

No, no confirmation so far, that's why we need more to investigate

peo

@olivierlambert said in Continuous Replication jobs creates full backups every time since 2025-09-06 (xo from source):

No, no confirmation so far, that's why we need more to investigate

Great (or maybe not) news.. At least on my system this is very easy to replicate.. On a tiny test machine that only have had a normal backup before, first time doing replication copy (to H2 SSD) it sends it as full as expected. Just did some simple operations on the machine (like apt update, no upgrade), then retried:

peo

Too bad it becomes that unexplainable slow when it has to fall back to a full backup:

The first (full) backup was 10x faster:

Andrew

@olivierlambert This happens to me too with a459015ca91c159123bb682f16237b4371a312a6.

I did open an issue https://github.com/vatesfr/xen-orchestra/issues/8969

andrew64k created this issue in vatesfr/xen-orchestra

open a459015 causes delta replication to always be full #8969

olivierlambert

@Andrew And it doesn't with the commit just before?

Andrew

@olivierlambert Correct. Running commit 4944ea902ff19f172b1b86ec96ad989e322bec2c works.

olivierlambert

@florent so it's like https://github.com/vatesfr/xen-orchestra/commit/a459015ca91c159123bb682f16237b4371a312a6 might introduced a regression?

0

fbeauchamp committed to vatesfr/xen-orchestra

Fix(replication): VDI_NOT_MANAGED error  (#8935)

from ticket #40151

florent

@Andrew then again, with such a precise report, the fix is easier

the fix should join master soon

olivierlambert

If you want to test, you can switch to the relevant branch, which is fix_replication

https://github.com/vatesfr/xen-orchestra/pull/8971

fbeauchamp opened this pull request in vatesfr/xen-orchestra

closed fix(backups): replication doing always a full #8971

Andrew

@olivierlambert fix_replication works for some... but most of my VMs now have an error:
the writer IncrementalXapiWriter has failed the step writer.checkBaseVdis() with error Cannot read properties of undefined (reading 'managed').

      "result": {
        "message": "Cannot read properties of undefined (reading 'managed')",
        "name": "TypeError",
        "stack": "TypeError: Cannot read properties of undefined (reading 'managed')\n    at file:///opt/xo/xo-builds/xen-orchestra-202509150025/@xen-orchestra/backups/_runners/_writers/IncrementalXapiWriter.mjs:29:15\n    at Array.filter (<anonymous>)\n    at IncrementalXapiWriter.checkBaseVdis (file:///opt/xo/xo-builds/xen-orchestra-202509150025/@xen-orchestra/backups/_runners/_writers/IncrementalXapiWriter.mjs:26:8)\n    at file:///opt/xo/xo-builds/xen-orchestra-202509150025/@xen-orchestra/backups/_runners/_vmRunners/IncrementalXapi.mjs:159:54\n    at callWriter (file:///opt/xo/xo-builds/xen-orchestra-202509150025/@xen-orchestra/backups/_runners/_vmRunners/_Abstract.mjs:33:15)\n    at IncrementalXapiVmBackupRunner._callWriters (file:///opt/xo/xo-builds/xen-orchestra-202509150025/@xen-orchestra/backups/_runners/_vmRunners/_Abstract.mjs:52:14)\n    at IncrementalXapiVmBackupRunner._selectBaseVm (file:///opt/xo/xo-builds/xen-orchestra-202509150025/@xen-orchestra/backups/_runners/_vmRunners/IncrementalXapi.mjs:158:16)\n    at process.processTicksAndRejections (node:internal/process/task_queues:105:5)\n    at async IncrementalXapiVmBackupRunner.run (file:///opt/xo/xo-builds/xen-orchestra-202509150025/@xen-orchestra/backups/_runners/_vmRunners/_AbstractXapi.mjs:378:5)\n    at async file:///opt/xo/xo-builds/xen-orchestra-202509150025/@xen-orchestra/backups/_runners/VmsXapi.mjs:166:38"
      }

olivierlambert

Feedback for you @florent

JB

@peo Yes, same problem!

Andrew

@olivierlambert @florent Looks like 1471ab0c7c79fa6dca9a1598e7be2a141753ba91 (in current master) has fixed the issue.

But, during testing I found a new related issue (no error messages). Running current XO master b16d5...

During the normal CR delta backup job, two VMs shows the warning message: Backup fell back to a full but it did a delta (not full) backup.

I caused this by running a different CR backup job (for testing) that used the two same VMs but to a different SR. The new test job did a full backup and delta backup correctly. Since there is only one CBT snapshot the normal backup job did recognize that the snapshot was not for the original normal backup job and should have done a full (and started a new CBT snapshot), but it only did a short delta.

As the VMs were off, I guess it could correctly use the other CBT snapshot as no blocks had changed... Just odd that CR backup said it was going to do a full and then did a delta.

JB

@Andrew

florent

@JB this means that it should have done a delta ( as per the full backup interval ) , but had to fall back to a full for at least one disk
this can happens after a failed transfer, a new disk added and some edge case. This issue was not really visible before the latest release, even if the impacts can be important , saturating network and storage.

We are investigating this ( especially @Bastien-Nollet ) , and expect to have a fix and/or an explanation fast

Are you using "purge snapshot data" option ? Are there anything on the journalctl logs ?

JB

@florent Thanks, Florent! No, No!