System behaviour when evacuating REP 1 objects #754

New issue

Open

opened 2025-12-28 17:20:36 +00:00 by sami · 3 comments

sami commented

2025-12-28 17:20:36 +00:00

Owner

Originally created by @fyrchik on GitHub (Sep 23, 2022).

We want to behave correctly when evacuating objects stored only on a single node.
Here are 2 situations:

Move node to maintenance -> evacuate objects
Evacuate objects -> move node to maintenance

In (1) the objects could be unavailable between the first and the second step.
In (2) the objects could become unavailable if policer on another node removes them right after the evacuation.

So we need to take special care here. The proposal is following:

Allow node to answer some requests in maintenance mode if possible.
Change its state in the netmap.
Make sure policer takes netmap into account even if a node answers requests: if an object is stored on a node that is under maintenance in a single replica, make sure it is replicated.

Originally created by @fyrchik on GitHub (Sep 23, 2022). We want to behave correctly when evacuating objects stored only on a single node. Here are 2 situations: 1. Move node to maintenance -> evacuate objects 2. Evacuate objects -> move node to maintenance In (1) the objects could be unavailable between the first and the second step. In (2) the objects could become unavailable if policer on another node removes them right after the evacuation. So we need to take special care here. The proposal is following: 1. Allow node to answer some requests in maintenance mode if possible. 2. Change its state in the netmap. 3. Make sure policer takes netmap into account even if a node answers requests: if an object is stored on a node that is under maintenance in a single replica, make sure it is replicated.

sami added the

bug

labels

2025-12-28 17:20:36 +00:00

sami commented

2025-12-28 17:20:36 +00:00

Author

Owner

@fyrchik commented on GitHub (Sep 23, 2022):

Actually, this may be an SDK issue -- we could extend replica selection algorithm to take state into account during HRW sorting.

@fyrchik commented on GitHub (Sep 23, 2022): Actually, this may be an SDK issue -- we could extend replica selection algorithm to take state into account during HRW sorting.

sami commented

2025-12-28 17:20:36 +00:00

Author

Owner

@fyrchik commented on GitHub (Oct 25, 2022):

Let's make it a bug -- we do not want policer to remove a single available copy.
However, policer still needs to avoid creating additional copies.

@fyrchik commented on GitHub (Oct 25, 2022): Let's make it a bug -- we do not want policer to remove a single available copy. However, policer still needs to avoid _creating_ additional copies.

sami commented

2025-12-28 17:20:36 +00:00

Author

Owner

@carpawell commented on GitHub (Nov 14, 2022):

@fyrchik, was it closed via the #1987?

@carpawell commented on GitHub (Nov 14, 2022): @fyrchik, was it closed via the #1987?