Skip to content
TensorDB

Scaling

Current: Vertical Scaling

TensorDB scales vertically through:

  • More shards: Increase shard_count for more write parallelism
  • More memory: Larger caches and memtables reduce I/O
  • Faster storage: NVMe SSDs minimize read latency
  • More cores: Each shard can use its own core

Planned: Horizontal Scaling

Shard Distribution

Distribute shards across multiple nodes:

Node 1: Shards 0-3
Node 2: Shards 4-7
Node 3: Shards 8-11

Resharding

When adding nodes, shards can be migrated:

  1. Create new shards on the new node
  2. Stream data from source shards
  3. Atomically switch routing
  4. Clean up source shards

Cross-Node Queries

SQL queries that span multiple shards will be distributed:

Client → Coordinator → Fan-out to shard nodes → Merge results → Response

Performance Expectations

ConfigurationPoint ReadPoint WriteThroughput
1 node, 4 shards276ns1.9µs~500K writes/s
3 nodes, 12 shards~500ns~3µs~1.5M writes/s
5 nodes, 20 shards~600ns~4µs~2.5M writes/s