== Repo setup
./setup.sh
== Starting daemons:
influxdb
./sw/influx/usr/bin/influxd
victoriametrics
./sw/victoriametrics/victoria-metrics-prod \
-retentionPeriod 1000 \
-search.disableCache \
-search.maxConcurrentRequests=$[$(nproc) * 2] \
-search.maxQueryDuration=60s
kudu
# Disable replication since we'll only run one tserver.
kudu master run -fs-wal-dir /data/m -default-num-replicas 1 &
# Only retain a few seconds of history to save disk space and IO.
# The other comparable systems don't allow time-travel or differential
# backups.
kudu tserver run -fs-wal-dir /data/ts --tablet_history_max_age_sec=5 &
# We use large batches on load, so allow the web server to accept them
kudu-tsdb -logtostderr -webserver_max_post_length_bytes=100000000
Clickhouse
Install ClickHouse from APT repos as described on appropriate documentation. Run it with the default configuration.
== Running benchmarks
Start the appropriate server per above, and then run the following with one of
influx, kudu, clickhouse, or victoriametrics:
# Set the number of concurrent clients as you like. Twice the number of vCPUs
# seems to maximize throughput for Kudu but has some throughput collapse issues
# on VictoriaMetrics.
NUM_CLIENTS=$[$(nproc)*2]
./env/bin/python -u benchmark.py load influx
./env/bin/python -u benchmark.py run-queries influx --workers=$NUM_CLIENTS
Various results will be emitted into logs/
== Summarizing results
Simple tabulation can be done on the command line:
cat logs/*json | jq '[.workers, .workload, .system, .qps] | @tsv' -r | sort -n | column -t
or use the included jupyter notebook graphs.ipynb to generate nice graphs.