Cockroach Labs Estate economics What idle nodes, provisioned bytes and peak-sized bills cost, worked out from your estate.
  1. Nodes 15 baseline
  2. Avg CPU utilization 15.0% idle hardware
  3. Storage provisioned 300 GB 50 GB of data
  4. Est. monthly saving 0%

Step 6 of 8 - Storage

Losing a storage server, and getting it back

Kill a blob server, then bring it back. The error count doesn't move either way, and when you bring it back decides whether the failure cost anything at all.

Kill a blob server Server 1 Server 2 Server 3 Server 4 Server 5 Server 6
Blob serversblob-1483 objectsblob-2483 objectsblob-3483 objectsblob-4483 objectsblob-5483 objectsblob-6483 objectsS3 backstop copy - 11-nines durabilityBlob server topologyAll blob servers healthy. Click one to kill it.

t = 0 s healthy time-lapse: 1 s on screen = 2 min of cluster time

All six servers are heartbeating.

Under-replicated objects
0
Repaired
0 (0%)
Copied
0 GB of 50 GB
Reads served from object storage
0
Errors
0
Read p99
0.1 ms Illustrative, not measured.
Data on one serverthe estate's 300 GB of local copies across 6 servers50 GB
Objects50 GB at an average 106 MB per object483
Time to suspectblob.suspect_after: replicas stop counting; with the archive valid, nothing is replaced30 s
Time to failedblob.failed_after: only now does up-replication start5 min
Copy raterepair.op_rate_limit 256 MiB/s x repair.concurrency 8 x repair.max_controllers 24096 MiB/s
Time to repair50 GB at 4096 MiB/s, once the disk is failed12 s
  1. Click a blob server to kill it.

S3 serves the reads while the replica rebuilds

S3 holds a copy, so reads fall back to it while the replica rebuilds.

How this is worked out
  • This scene is a time-lapse and it's the only one on the site that is. The controller waits 5 min without a heartbeat before it will replace a replica, so a 1x drill would be five minutes of nothing. One second on screen is 2 min of cluster time, and the clock in the readout carries the real figure.
  • Every duration here comes from a rate or a threshold the cluster actually ships with: blob.suspect_after, blob.failed_after, repair.op_rate_limit x repair.concurrency x repair.max_controllers for the copy, and gc.zombie_min_age with gc.batch_rate for the cleanup. The rate is a ceiling the settings impose, so 12 s for 50 GB is the floor, not the forecast.
  • Bringing the server back before 5 min costs nothing at all: a suspect disk keeps its replicas while its objects still hold a valid archive copy and aren't down a second replica, because up-replicating a disk that's about to return is work the return immediately undoes. Fall below either of those and the controller stops waiting. Come back later and the repair has already happened - the returning copies are surplus, and clearing them up is background GC that no read is waiting on.