Logs
Every log of the fleet in one place, searched in one line, parsed and masked on arrival, with control over their volume.
Where logs come from
- Linux: the systemd journal and the files declared in
logs.files. - Windows: event logs (System and Application by default; other channels and filters: see Windows).
- Applications: OpenTelemetry (
/v1/logs), directly or through the agent's relay. - A JSON line is decoded by the agent:
messageormsg,levelorseverity, the rest as attributes.
See log configuration.
What the video shows
- The Logs page and its histogram by status.
- The query status:error OR status:critical filters the lines; facets update.
- A line in detail: attributes, tags, context.
- The Patterns tab groups similar messages and counts them.
Search
| Syntax | Effect |
|---|---|
timeout connection | lines containing both words (implicit AND, case-insensitive) |
"exact phrase" | the phrase as is |
service:nginx, host:web-01, status:error, source:journald | field equality; service:shop* for a prefix |
@http.status_code:502 | attribute value |
-service:sshd | exclusion |
status:error OR status:critical | either one |
Statuses: emergency, alert, critical, error, warning, notice, info, debug.
The page shows the histogram by status and the facets: status, service, host, source, and, when present, HTTP code and method. Clicking a value adds it to the query.
- Live: new lines arrive every 2 seconds.
- Context: the lines of the same host and service around a line.
- Patterns: similar messages grouped into templates (numbers, identifiers, addresses, paths replaced), counted, with their share and statuses. A click filters on the pattern.
- Views: a search can be saved and found again in the Views menu.
- Explain: from a line's details, S33D's AI describes what is happening (see Explain and repair).
Parsing on arrival
HYPHA parses common formats into attributes, without overwriting an existing attribute:
- uvicorn, nginx and Apache access logs (combined format), HAProxy (httplog):
http.method,http.url_path,http.status_code,http.useragent,network.client.ip,network.bytes_written,duration,haproxy.frontend,haproxy.backend,haproxy.server; sshd,sudo,su,fail2ban:evt.name,evt.outcome,usr.name,network.client.ip,sudo.command,fail2ban.jail.
An info line with an HTTP code of 500 or more becomes an error, 400 or more a warning. A more severe status already set is never lowered.
Exclusion filters and sampling
Access logs and health probes often make up most of the volume. Settings › Logs shows volume by service and by host, and sets filters applied on arrival, before writing:
- Exclude: lines matching the query are dropped.
- Sample: a share of the lines is kept (for example 10%), drawn in a stable way.
- The first matching filter decides; reorder by dragging.
- Before saving, the preview gives the lines per hour concerned, their share of the total and five examples.
- Errors and more severe lines are never dropped by sampling, only by an explicit exclusion on the status.
From the Logs page, Exclude these lines pre-fills a filter with the current query. Dropped lines do not count in your usage.
Masking sensitive data
Applied on arrival, before writing, to the message and text attributes. Masked data becomes [masqué:<rule>].
| Rule | What it masks | State in a new workspace |
|---|---|---|
secrets | Bearer tokens, password=…, AWS, GitHub and Slack keys, HYPHA keys, JWTs, PEM private keys | on |
credit_card | card numbers (Luhn check) | offered |
iban | IBANs | offered |
email | email addresses | offered |
phone_fr | French phone numbers | offered |
ipv4_public | public IPv4 addresses | offered |
fr_nir | French social security numbers | offered |
An administrator turns rules on and tests a sample text in Settings › Logs.
Log-based metrics
A log query can feed a metric: a count of lines, or the distribution of an attribute (.avg, .p50, .p95, .p99, .max), by group. These metrics are plotted and alerted on like any other and remain after the logs' retention ends.
Application errors
An error line carrying a stack trace (Python, Java, Go, Node) or error.kind and error.stack attributes feeds the Errors page, with errors from traces and browsers. See Traces.
Retention
Retention depends on the plan: see pricing. When self-hosting, you set it.
Updated: 19 November 2026