Tell now ingests infrastructure metrics as a first-class data type. CPU usage, memory trends, network throughput, request latency distributions — stored in ClickHouse and queryable with the same tools you use for product analytics.

Collect with Witness, query in Tell. Your server metrics live next to your product events in the same database.

Three metric types

Gauge — point-in-time values. CPU percentage, available memory, disk space, queue depth
Counter — cumulative or delta counts. Bytes sent, requests served, errors total. Tell handles both Prometheus-style cumulative and StatsD-style delta temporality
Histogram — full distributions with bucket boundaries, count, sum, min, max. Request latency at p50, p95, p99 without pre-aggregation

Every metric carries string and integer labels for slicing by host, service, core, device, interface, status code — any dimension you need.

Query and breakdown

Eleven aggregation functions, all with time bucketing and label breakdown:

avg, sum, min, max — standard aggregates across any time window
last — most recent observed value, useful for gauge snapshots
rate — change over period, for counters that report cumulative totals
count — number of data points in the window
p50, p75, p90, p95, p99 — percentile queries directly on stored values

Break down any metric by any label. CPU percentage by core, disk I/O by device, network bytes by interface, request latency by endpoint — one query, one chart.

Witness integration

The recommended way to collect infrastructure metrics. Install the Witness agent on your servers and metrics flow automatically — CPU, memory, disk, network, load, TCP state, containers, and top processes out of the box.

Same connection — metrics and logs ship over a single TCP connection
Hourly checkpoints — cumulative totals alongside deltas, recovers exact counts after outages
15-second resolution — configurable down to sub-second intervals

Any Tell SDK can also send custom application metrics using the same gauge, counter, and histogram types.

Also in this update

Delta and cumulative temporality — works with both Prometheus-style and StatsD-style instrumentation
Integer labels for numeric dimensions like status codes and ports — efficient range queries without string conversion
Session correlation for client-side metrics — tie app performance to user sessions
90-day retention with daily partitioning, configurable per workspace