Skip to content

Latest commit

 

History

History
40 lines (32 loc) · 2.38 KB

faq.md

File metadata and controls

40 lines (32 loc) · 2.38 KB

FAQ & troubleshooting

Start-up

Symptom Cause Fix
Pool overlaps with other one on this address space 10.203.0.0/24 is taken VORTEX_SUBNET_PREFIX=10.77.0 docker compose up -d
port is already allocated 9080 or 9081 in use VORTEX_GATEWAY_PORT=9180 VORTEX_DASHBOARD_PORT=9181 docker compose up -d
Swagger UI 404, /tsd works Nodes still starting (20–30 s); the gateway sends everything to the console until they are healthy Wait for (healthy) in docker compose ps
container name … is already in use Another stack uses the same names VORTEX_CONTAINER_PREFIX=vortex2
pull access denied / 401 Stale registry login on this machine docker logout (images are public), or log in again

Redis

Question Answer
Do I need Redis? No. Without it, keys beyond VORTEX_CACHE_MAX_KEYS per node are deleted, oldest first
Redis on my laptop is not reached Use host.docker.internal; on Linux Engine also uncomment extra_hosts in the compose file
Redis goes down Nodes keep serving from memory; spills fail until it is back. The start-up log warns if it is unreachable
Why is a query slow? It read keys that moved to Redis: one round trip each. Raise VORTEX_CACHE_MAX_KEYS

Data and failures

Question Answer
Is data kept across restarts? Yes: snapshot at stop and every 5 min, in each node's volume. docker compose down -v deletes it
What if every node is killed at once? Back to the last snapshot: up to VORTEX_SNAPSHOT_INTERVAL lost
What happens when the leader dies? A new leader in about 5 s. Writes in flight fail (HTTP 503/502); a few may still have been stored
Should clients retry failed writes? Only if a duplicate sample is acceptable
Can I add nodes to a running cluster? Membership is fixed at start: add the node to VORTEX_CLUSTER_PEERS and recreate the stack
Why don't writes get faster with more nodes? Every write is applied on the leader; nodes add read capacity and redundancy

Queries

Question Answer
step rejected It must divide an hour (1–30) or a day (60–720)
Points don't line up with my local hour Pass z, e.g. z=Asia/Shanghai
An empty bucket shows null No samples in that minute; count is 0
Is there a query language? No, by design: range, step, window, z