layout doc: write explanations for bizarre scenarios

This commit is contained in:
Alex Auvolat 2023-09-22 16:07:46 +02:00
parent 405aa42b7d
commit 8d07888fa2

View File

@ -93,9 +93,23 @@ follow the following recommendations:
## Understanding unexpected layout calculations
When adding, removing or modifying nodes in a cluster layout, sometimes
unexpected assigntations of partitions to node can occure. These assignations
are in fact normal and logical, given the objectives of the algorihtm. Indeed,
**the layout algorithm prioritizes moving less data between nodes over the fact
of achieving equal distribution of load**. This section presents two examples
and illustrates how one can control Garage's behavior to obtain the desired
results.
### Example 1
In this example, a cluster is originally composed of 3 nodes in 3 different
zones (data centers). The three nodes are of equal capacity, therefore they
are all fully exploited and all store a copy of all of the data in the cluster.
Then, a fourth node of the same size is added in the datacenter `dc1`.
As illustrated by the following, **Garage will by default not store any data on the new node**:
```
$ garage layout show
==== CURRENT CLUSTER LAYOUT ====
@ -146,8 +160,37 @@ dc3 Tags Partitions Capacity Usable capacity
TOTAL 256 (256 unique) 1000.0 MB 1000.0 MB (100.0%)
```
While unexpected, this is logical because of the following facts:
- storing some data on the new node does not help increase the total quantity
of data that can be stored on the cluster, as the two other zones (`dc2` and
`dc3`) still need to store a full copy of everything, and their capacity is
still the same;
- there is therefore no need to move any data on the new node as this would be pointless;
- moving data to the new node has a cost which the algorithm decides to not pay if not necessary.
This distribution of data can however not be what the administrator wanted: if
they added a new node to `dc1`, it might be because the existing node is too
slow, and they wish to divide its load by half. In that case, what they need to
do to force Garage to distribute the data between the two nodes is to attribute
only half of the capacity to each node in `dc1` (in our example, 500M instead of 1G).
In that case, Garage would determine that to be able to store 1G in total, it
would need to store 500M on the old node and 500M on the added one.
### Example 2
The following example is a slightly different scenario, where `dc1` had two
nodes that were used at 50%, and `dc2` and `dc3` each have one node that is
100% used. All node capacities are the same.
Then, a node from `dc1` is moved into `dc3`. One could expect that the roles of
`dc1` and `dc3` would simply be swapped: the remaining node in `dc1` would be
used at 100%, and the two nodes now in `dc3` would be used at 50%. Instead,
this happens:
```
==== CURRENT CLUSTER LAYOUT ====
ID Tags Zone Capacity Usable capacity
@ -197,3 +240,34 @@ dc3 Tags Partitions Capacity Usable capacity
a11c7cf18af29737 node4 63 (0 new) 1000.0 MB 246.1 MB (24.6%)
TOTAL 256 (256 unique) 2.0 GB 1000.0 MB (50.0%)
```
As we can see, the node that was moved to `dc3` (node4) is only used at 25% (approximatively),
whereas the node that was already in `dc3` (node3) is used at 75%.
This can be explained by the following:
- node1 will now be the only node remaining in `dc1`, thus it has to store all
of the data in the cluster. Since it was storing only half of it before, it has
to retrieve the other half from other nodes in the cluster.
- The data which it does not have is entirely stored by the other node that was
in `dc1` and that is now in `dc3` (node4). There is also a copy of it on node2
and node3 since both these nodes have a copy of everything.
- node3 and node4 are the two nodes that will now be in a datacenter that is
under-utilized (`dc3`), this means that those are the two candidates from which
data can be removed to be moved to node1.
- Garage will move data in equal proportions from all possible sources, in this
case it means that it will tranfer 25% of the entire data set from node3 to
node1 and another 25% from node4 to node1.
This explains why node3 ends with 75% utilization (100% from before minus 25%
that is moved to node1), and node4 ends with 25% (50% from before minus 25%
that is moved to node1).
This illustrates another principle of the layout computation: **if there is a
choice in moving data out of some nodes, then all links between pairs of nodes
are used in equal proportions** (this is approximately true, there is
randomness in the algorihtm to achieve this so there might be some small
fluctuations, as we see above).