[Question posted by a user on YugabyteDB Community Slack ]
Assuming we have multi-DC or a multi-zone aware replica placement with RF=3, each zone has 2 (az1) + 2 (az2) + 2 (az3) = 6 nodes . Each node has 1TB of space.
How much storage should we over-provision on each node ?
There are several points here:
YugabyteDB stores data compressed by default. We have seen 3-5x compression in a lot of workloads. In some workloads, if the app has already compressed large payload columns, then the additional compression that YugabyteDB can do is lesser. But assuming say replication factor is 3 and compression ratio is 4 – then if you have 5TB of user data-- you’ll need 5TB * 3 / 4 ==> 3.75TB of raw space across all your nodes.
There are essentially two key factors – replication factor and compression. You should plan for about 20% head room on each node to allow space for miscellaneous overheads (temporary additional space needed during compactions as well as metadata overheads). Application created table indexes also need space and must be planned/accounted for.
If one node is down in az1 for more than a certain time (by default 15 mins), then the only other candidate that can now host its data is the other remaining node in the same az. A more ideal setup to avoid this overhead would be to use smaller sized nodes but more of them. For example, 4 + 4 + 4. In this case, if one of the nodes is down, there are still 3 other nodes to redistribute the data over.
If p is percent utilization on every disk, then it has to have a buffer for p/(n-1) (in case there is a node down), where n is the number of nodes in the AZ/DC, then we recommend that p + p/(n-1) <= 80% to 85%. This is just good practice to not run the disk too close to edge of its available capacity. If we go with the 80% number, then simplifying above, p <= 80*(n-1)/n.
3 node per AZ, usable space = 53% (26.5% from other node + 20% spare)
4 node per AZ, usable space = 60% (~20% from other node + 20% spare)
6 node per AZ, usable space = 66.6% (~14% from other node + 20% spare)