Send back and force, large blocks?

[IvanVolosyuk]

I'm trying to re-sync my machine from backup, but in one of the datasets I have following issue. I can send data one way, but cannot send the same data another way. Before the backups became out of sync due to operator error it was working fine and the snapshot I deleted just for testing was also received correctly previously. Any ideas what can be so special about this dataset and how can I restore from backup and keep sending to the same backup. Features of both pools are exactly the same and have support for large blocks.

$ zfs list -t snapshot archive/replica/home/ivan-cache
NAME                                         USED  AVAIL  REFER  MOUNTPOINT
archive/replica/home/ivan-cache@2025-07      937M      -  21.2G  -
archive/replica/home/ivan-cache@2025-08      895M      -  21.7G  -
archive/replica/home/ivan-cache@2025-09      876M      -  21.8G  -
archive/replica/home/ivan-cache@2025-10      358M      -  21.7G  -
archive/replica/home/ivan-cache@2025-10-04   148M      -  21.8G  -
archive/replica/home/ivan-cache@2025-10-05  82.5M      -  21.8G  -
archive/replica/home/ivan-cache@2025-10-06  77.7M      -  21.8G  -
archive/replica/home/ivan-cache@2025-10-07     0B      -  21.7G  -
$ zfs send -L -R archive/replica/home/ivan-cache@2025-10-07 | zfs recv ssd/home/ivan-cache2 -x compression
$ zfs list -t snapshot ssd/home/ivan-cache2
NAME                              USED  AVAIL  REFER  MOUNTPOINT
ssd/home/ivan-cache2@2025-07     1.12G      -  23.0G  -
ssd/home/ivan-cache2@2025-08     1.10G      -  23.7G  -
ssd/home/ivan-cache2@2025-09     1.07G      -  23.8G  -
ssd/home/ivan-cache2@2025-10      471M      -  23.7G  -
ssd/home/ivan-cache2@2025-10-04   187M      -  23.8G  -
ssd/home/ivan-cache2@2025-10-05  97.2M      -  23.8G  -
ssd/home/ivan-cache2@2025-10-06  88.6M      -  23.8G  -
ssd/home/ivan-cache2@2025-10-07     0B      -  23.7G  -
$ zfs rollback -r archive/replica/home/ivan-cache@2025-10-06
$ zfs send -L -i ssd/home/ivan-cache2@2025-10-06 ssd/home/ivan-cache2@2025-10-07 | zfs recv archive/replica/home/ivan-cache
cannot receive incremental stream: incremental send stream requires -L (--large-block), to match previous receive.
$ zfs send  -i ssd/home/ivan-cache2@2025-10-06 ssd/home/ivan-cache2@2025-10-07 | zfs recv archive/replica/home/ivan-cache
cannot receive incremental stream: incremental send stream requires -L (--large-block), to match previous receive.

I think the problem can be that there was not a single file with large blocks in archive/replica/home/ivan-cache as the original source dataset had recordsize=128k or even less. The archive/replica/home/ivan-cache has an inherited recordsize=1M. When the data was sent to ssd/home/ivan-cache2 the recordsize on the destination somehow was 128k (inherited?, shouldn't it be 1M? It wasn't explicitly set on source archive/replica/home/ivan-cache, but it was inherited from archive/replica).

I solved it doing several changes at once, so I don't know which one helped:

• Set recordsize=1M on restored source ssd/home/ivan-cache2
• Created a file with large blocks there as well, dd if=/dev/urandom of=/ssd/home/ivan-cache2 bs=1M count=2
• Deleted previous snapshots which didn't exist at destination (ssd/home/ivan-cache2@2025-10-07)
• Created a new snapshot, which now included at least one file with large blocks.
• Sending successfully to replica with -L.

I still don't understand what exactly happened. Should the restored source have to be tainted with a large block file to be considered worthy candidate to send with -L if destination doesn't have any large block files, but had it previously at some point?