Running lightweight models on the edge gateway: latency vs privacy阅读 2.8k
Move inference from the cloud to the gateway so sensor data can stay on-site for anomaly detection and simple decisions.
Blog
Full list of all articles; the home page shows paginated previews (20 per page). Permalink /blog.
Move inference from the cloud to the gateway so sensor data can stay on-site for anomaly detection and simple decisions.
Ship updates without bricking devices—partitioning, dual-bank images, and canary groups.
Bridge plant floors to telemetry buses without melting your broker or historians.
Plan capacity before you deploy hundreds of nodes in one yard.
From a single project to multi-tenant setups—how to name topics and model shadow state without painting yourself into a corner.
Keep scans predictable and map registers to a clean thing model upstream.
Start with assets, alarms, and throughput—not every bolt on day one.
Seasonality and maintenance windows will fool naive thresholds—plan for them.
Vectorize metric summaries and alert text so you can query historical conditions and runbooks in conversation.
Match orchestration weight to power, staff skills, and offline tolerance.
Keep query latency sane while preserving evidence for audits.
Design alert delivery so duplicate bursts do not page the whole team.
Pick the contract style that matches your connectivity and client stacks.
Reduce false positives when doors open or trucks reload.
Ship useful vision features without storing more pixels than you need.
Treat the PLC as the source of truth for physical state.
Separate namespaces early so one customer cannot enumerate another’s fleet.
Use version vectors or last-writer-wins with explicit domains.
Normalize units and clock skew before comparing sites.
Without failure history, start with health scores—not magic failure dates.
Model dependencies for blast-radius during outages and upgrades.
Match local storage to query patterns and crash safety requirements.
Assume compromise; limit lateral movement from cheap cameras and sensors.
Not legal advice—just engineering habits that help audits go smoother.
Tie production events to telemetry for faster root cause.
Operators need clarity when the uplink dies.
Authenticate actions and avoid broadcasting sensitive readouts.
Replay shapes and edge cases your real fleet rarely hits.
Price SKUs with real marginal cost—not flat per-device guesses.
Retire hardware without surprise downtime during busy seasons.
Tune scan rates so panels stay responsive while historians stay fresh.
Give operators enough context to trust—or override—local inference.
Match delivery semantics to billing, safety, and deduplication needs.
Avoid blocking the PLC driver thread with slow database writes.
Replace shared VPN passwords with audited, time-bound sessions.
Bad timestamps break correlation across sites—plan for drift.
Pull times and disk matter when uplinks are megabits, not gigabits.
If nobody trusts alerts, automation will be ignored—quantify the pain.
Plan factory resets and re-keying without truck rolls.
Sometimes a 5-minute tumble window beats a custom state machine.
Never bypass the BMS safety chain from the cloud API.
Retrain triggers should fire on precision drops, not calendar dates.
Dual gateways need deterministic merge rules for command topics.
Freeze schemas for compliance windows; document transformations.
Stagger renewals and support two valid keys during overlap.
Retail and ops boards should degrade visibly, not go blank.
Pair point sensors with zone logic to survive mopping and condensation.
Keep error traces rich while summarizing healthy device chatter.
Site surveys beat guesswork when metal racks move seasonally.
Load test with burst shapes, not flatlines.