serving-only.toml

[storage]
# Shared buffer is useless to a serving-only compute node.
shared_buffer_capacity_mb = 1
# Compactor is irrelevant to a serving-only compute node.
compactor_memory_limit_mb = 1
# Suppose a serving-only CN has 16GB memory in total, we recommend using 30% by default: 16*0.3=4.8
block_cache_capacity_mb = 4800
# Suppose a serving-only CN has 16GB memory in toal, we recommend using 10% by default: 16*0.1=1.6
meta_cache_capacity_mb = 1600
# There will be 16GB - 4.8GB - 1.6GB - reserved memory 16GB * 0.2 (recommended and hard coded by developers) = 6.4GB for executing serving queries. We call it "compute memory".

# If one batch query is resource-intensive and its runtime memory consumption exceeds 6.4GB, it will kill itself automatically before triggering out-of-memory error. The reserved memory is allocated for all types of temporary memory usage and slackness.
# Although we don't recommend users to execute OLAP-style batch queries that take a huge amount of input data, e.g. more than dozens of GB, users can adjust the configuration if such batch query is needed and the default configuration above leaves too few "compute memory".
# Feel free to allocate less memory for "block cache" and "meta cache" to increase "compute memory".