mirror of
https://github.com/zrepl/zrepl.git
synced 2024-11-21 16:03:32 +01:00
37 lines
2.2 KiB
ReStructuredText
37 lines
2.2 KiB
ReStructuredText
|
|
Migrating Sending Side
|
|
~~~~~~~~~~~~~~~~~~~~~~
|
|
|
|
**Objective**:
|
|
Move sending-side zpool to new hardware.
|
|
Make the move fully transparent to the sending-side jobs.
|
|
After the move is done, all sending-side zrepl jobs should continue to work as if the move had not happened.
|
|
In particular, incremental replication should be able to pick up where it left before the move.
|
|
|
|
Suppose we want to migrate all data from one zpool ``oldpool`` to another zpool ``newpool``.
|
|
A possible reason might be that we want to change RAID levels, ``ashift``, or just migrate over to next-gen hardware.
|
|
|
|
If the pool names are different, zrepl's matching between sender and receiver dataset will break becase the receive-side dataset names contain ``oldpool``.
|
|
To avoid this, we will need the name of the new pool to match that of the old pool.
|
|
The following steps will accomplish this:
|
|
|
|
1. Stop zrepl.
|
|
2. Create the new pool: ``zpool create newpool ...``
|
|
3. Take a snapshot of the old pool so that you have something that you can ``zfs send``.
|
|
For example, run ``zfs snapshot -r oldpool@migration_oldpool_newpool``.
|
|
4. Send all of the oldpool's datasets to the new pool:
|
|
``zfs send -R oldpool@migration_oldpool_newpool | zfs recv -F newpool``
|
|
5. Export the old pool: ``zpool export oldpool``
|
|
6. Export the new pool: ``zpool export newpool``
|
|
7. (Optional) Change the name of the old pool to something that does not conflict with the new pool.
|
|
We are going to use the name ``oldoldpool`` in this example.
|
|
Use ``zpool import`` with no arguments to see the pool id.
|
|
Then ``zpool import <id> oldoldpool && zpool export oldoldpool``.
|
|
8. Import the new pool, while changing the name to match the old pool: ``zpool import newpool oldpool``
|
|
9. Start zrepl again and wake up the relevant jobs.
|
|
10. Use ``zrepl status`` or you monitoring to ensure that replication works.
|
|
The best test is an end-to-end test where you write some junk data on a sender dataset and wait until a snapshot with that data appears on the receiving side.
|
|
11. Once you are confident that replication is working, you may dispose of the old pool.
|
|
|
|
Note that, depending on pruning rules, it will not be possible to switch back to the old pool seamlessly, i.e., without a full re-replication.
|