{{ summaryTitle }} {{ summaryValue }} {{ summaryLine }} {{ badge.label }} {{ badge.value }}
Name the on-call stream represented by this sample.
Compare the classified sample with your review ceiling.
%
Paste one CSV sample, drop one text file, or browse locally. Nothing is uploaded.
{{ sourceMeta }}
{{ sourceStatus || 'CSV parsing and calculation stay in this browser.' }}
Prioritize recurring alert-and-route groups rather than one-off rows.
alerts
Use zero when acknowledgement timing is not part of this review.
min
Choose from 1 to 25 routes; all alert rows remain available in the evidence ledger.
routes
Use detailed precision when reviewing small samples or closely spaced route rates.
MeasureValueReview noteCopy
{{ row.measure }}{{ row.value }}{{ row.note }}

{{ tableExportAnnouncement }}

Alert groupRouteAlertsNoise rateRecommended reviewCopy
No repeated noisy or unclassified groups meet the current threshold.
{{ row.alert }}{{ row.route }}{{ row.total }}{{ formatPercent(row.noise_rate, percentDigits) }}{{ row.action }}

{{ tableExportAnnouncement }}

AlertRouteDispositionClassAck delayCopy
{{ row.alert }}{{ row.route }}{{ row.disposition || '—' }}{{ row.classification }}{{ row.ack_minutes === null ? '—' : `${formatNumber(row.ack_minutes, numberDigits)} min` }}

{{ tableExportAnnouncement }}

{{ chartExportAnnouncement }}

Error Budget Burn Rate Calculator Calculate SLO error-budget burn rate and compare page or ticket thresholds, then estimate runway from the budget that remains.
Queue Buffer Delay Calculator Calculate queueing delay from buffer depth and bottleneck rate, then compare occupancy and target buffer size with optional BDP context.
SLO Burn Rate Alerts Generator Generate multi-window Prometheus SLO burn-rate alerts with budget math and routing labels, then review thresholds before loading the YAML.
BGP AS Path Policy Checker Audit pasted BGP route paths for policy conflicts and trace each pass, watch or reject verdict before changing production router filters.

On-call attention is limited, so an alert earns its place only when it helps a responder decide or act. Repeated pages that need no human response compete with incidents that do. Over time, that weak signal-to-noise ratio can slow acknowledgement, hide recurring faults, and train responders to distrust the pager.

Alert noise rate expresses the non-actionable share of alerts whose outcomes have been reviewed. An alert that led to a page, incident, ticket, mitigation, or other human action counts as signal. A duplicate, false positive, flap, informational event, ignored event, or automatic recovery counts as noise. An uncertain outcome belongs in neither group until someone classifies it.

OutcomePractical meaningEffect on the rate
SignalA responder took a useful human action.Included in the denominator.
NoiseNo human action was needed, or the alert repeated an existing event.Included in both numerator and denominator.
UnclassifiedThe outcome is blank or does not match a recognized disposition.Excluded until reviewed.

A target is an operational policy, not a universal industry cutoff. Teams with different services, risk, traffic, and escalation paths may choose different ceilings. The useful comparison is a consistently classified sample against the same local policy, followed by review of the alert names and routes that contribute most of the noise.

Rate alone cannot show whether a quiet pager is healthy. Suppressing a real failure can improve the percentage while making detection worse. Keep actionable coverage intact, investigate repeated groups before changing thresholds, and compare similar time windows after routing or deduplication changes.

Acknowledgement time adds another clue, but it is not proof of alert quality. A slow median can reflect overload, poor routing, missing timestamps, or a difficult incident. Read it beside the classifications and the underlying alert evidence rather than treating it as a second noise score.

How to Use This Tool:

Start with one representative on-call stream and a review ceiling that your team already understands.

  1. Name the team or service and enter its Noise target. The target is the maximum acceptable percentage of classified alerts that required no human action.
  2. Paste or load the Alert sample CSV. A header is recommended; the expected concepts are alert name, created time, acknowledged time, disposition, route, and severity. Correct any empty or oversized sample before continuing.
  3. Set the Repeat group threshold to the smallest recurring alert-and-route count worth reviewing. Set the Acknowledgement target to zero when timing is outside this review.
  4. Read Noise review first, then use Tuning queue for recurring groups and Alert evidence to resolve unclassified rows before changing suppression or routing.

Interpreting Results:

The headline compares the calculated noise rate with the selected ceiling. A rate equal to the target is labeled at target; only a strictly larger rate is above target. The signal rate is the complementary share of the same classified denominator.

  • Resolve Unclassified rows before trusting a trend. They are excluded, so a large unresolved share can move the rate substantially.
  • Use Tuning queue to find repeated alert-and-route groups, but inspect the evidence before suppressing them. A mixed group may contain both duplicates and real incidents.
  • Remember that Routes above target counts only the busiest routes currently displayed. Increasing the route limit can reveal additional routes without changing the overall noise rate.
  • Compare median acknowledgement only when valid created-to-acknowledged timestamps exist and the same timing convention is used across samples.

Technical Details:

Classification comes before arithmetic. Recognized actionable dispositions become signal, recognized non-actionable dispositions become noise, and every other disposition remains unclassified. At least one signal or noise row is required because an all-unclassified sample has no valid denominator.

Formula Core

The percentage divides non-actionable alerts by all classified alerts. Unclassified rows do not enter either term.

Noise rate= NnoiseNsignal+Nnoise ×100%

Signal rate uses the same denominator with signal in the numerator. Percentages retain full precision internally and display one or two decimal places according to the selected precision.

Rule Core

RuleExact behavior
Signal vocabularyyes, true, 1, page, paged, incident, ticket, action, actionable, human, escalated, mitigated, fixed, or resolved-by-human.
Noise vocabularyno, false, 0, noise, no-action, false-positive, duplicate, flap, flapping, auto, auto-resolved, informational, info, suppressed, or ignored.
Target stateAbove when noise rate > target, at target when equal, otherwise below.
Repeated groupSame alert name and route, total count at or above the threshold, with at least one noisy or unclassified row.
Noisy displayed routeRoute noise rate > target among the displayed routes after ranking by total volume.
Acknowledgement medianMedian of valid nonnegative created-to-acknowledged intervals; missing or reversed timestamps are omitted.

CSV parsing accepts a header or a six-column positional layout. Common header aliases and disposition words are normalized for case, spaces, and underscores. Quoted cells and doubled quotation marks are handled, but the parser is intended for one compact operational sample rather than an unrestricted data export. The source is limited to 1 MiB.

The repeat threshold accepts 1 to 1,000 alerts, the acknowledgement target accepts 0 to 1,440 minutes, and the route display accepts 1 to 25 routes. Those bounds protect the review workflow; they do not define good alerting policy.

Privacy Notes:

Pasted text and locally selected CSV or text files are processed in the browser and are not uploaded. Alert samples can still contain service names, incident details, or internal routes, so remove sensitive fields before sharing exported tables or documents.

Worked Examples:

Repeated CPU alert above the team ceiling

Three rows for the same CPU alert and infrastructure route contain two non-actionable outcomes and one incident. The noise rate is 2 ÷ 3, or 66.67%. With a 40% target and a repeat threshold of three, the sample is above target and that alert-route pair enters the tuning queue. The incident row is the reason to review grouping and thresholds carefully instead of suppressing the alert outright.

References: