Apim Kql

APIM KQL Skill

Guide for creating Kusto Query Language (KQL) queries for Azure API Management monitoring data in Azure Monitor.

Quick Start

Basic Gateway Logs Query

ApiManagementGatewayLogs
| where TimeGenerated > ago(1h)
| project TimeGenerated, Method, Url, ResponseCode, TotalTime, BackendTime, CallerIpAddress
| order by TimeGenerated desc
| take 100

LLM Token Usage Query

ApiManagementGatewayLlmLog
| where TimeGenerated > ago(24h)
| summarize 
    TotalPromptTokens = sum(PromptTokens),
    TotalCompletionTokens = sum(CompletionTokens),
    TotalTokens = sum(TotalTokens)
by DeploymentName, ModelName

Tables Overview

For complete table schemas, see references/table-schemas.md.

Essential Query Patterns

Joining LLM Logs with Gateway Logs

let llmLogs = ApiManagementGatewayLlmLog 
| where DeploymentName != '';

let logsWithSubscription = llmLogs 
| join kind=leftouter ApiManagementGatewayLogs on CorrelationId 
| project 
    SubscriptionId = ApimSubscriptionId, 
    DeploymentName, 
    ModelName,
    PromptTokens, 
    CompletionTokens, 
    TotalTokens;

logsWithSubscription 
| summarize 
    SumPromptTokens = sum(PromptTokens), 
    SumCompletionTokens = sum(CompletionTokens), 
    SumTotalTokens = sum(TotalTokens) 
by SubscriptionId, DeploymentName

Error Analysis

ApiManagementGatewayLogs
| where TimeGenerated > ago(24h)
| where ResponseCode >= 400
| summarize ErrorCount = count() by ResponseCode, ApiId, LastErrorReason
| order by ErrorCount desc

Request Latency Analysis

ApiManagementGatewayLogs
| where TimeGenerated > ago(1h)
| summarize 
    AvgTotalTime = avg(TotalTime),
    AvgBackendTime = avg(BackendTime),
    P95TotalTime = percentile(TotalTime, 95),
    RequestCount = count()
by bin(TimeGenerated, 5m), ApiId
| order by TimeGenerated desc

Backend Performance

ApiManagementGatewayLogs
| where TimeGenerated > ago(1h)
| where BackendId != ""
| summarize 
    AvgBackendTime = avg(BackendTime),
    MaxBackendTime = max(BackendTime),
    ErrorCount = countif(BackendResponseCode >= 400),
    TotalRequests = count()
by BackendId
| order by AvgBackendTime desc

Token Usage by Time

ApiManagementGatewayLlmLog
| where TimeGenerated > ago(7d)
| summarize 
    TotalTokens = sum(TotalTokens),
    PromptTokens = sum(PromptTokens),
    CompletionTokens = sum(CompletionTokens)
by bin(TimeGenerated, 1h), DeploymentName
| order by TimeGenerated desc

MCP Tool Usage

ApiManagementGatewayMCPLog
| where TimeGenerated > ago(24h)
| summarize 
    CallCount = count(),
    UniqueClients = dcount(ClientName)
by ToolName, ServerName, Method
| order by CallCount desc

WebSocket Connection Events

ApiManagementWebSocketConnectionLogs
| where TimeGenerated > ago(1h)
| summarize 
    EventCount = count()
by EventName, Source, Destination
| order by EventCount desc

KQL Operators Reference

Filtering

| where TimeGenerated > ago(1h)
| where ResponseCode == 200
| where ApiId contains "openai"
| where isnotempty(BackendId)

Aggregation

| summarize count() by ApiId
| summarize avg(TotalTime), percentile(TotalTime, 95) by ApiId
| summarize sum(TotalTokens) by DeploymentName
| summarize dcount(CallerIpAddress) by bin(TimeGenerated, 1h)

Time Functions

| where TimeGenerated > ago(1h)
| where TimeGenerated between (datetime(2024-01-01) .. datetime(2024-01-31))
| where TimeGenerated >= startofmonth(now()) and TimeGenerated <= endofmonth(now())
| summarize ... by bin(TimeGenerated, 5m)

Joins

| join kind=leftouter OtherTable on CorrelationId
| join kind=inner OtherTable on $left.Field1 == $right.Field2

Projections

| project TimeGenerated, Method, Url, ResponseCode
| project-away _BilledSize, _IsBillable
| extend Duration = TotalTime - BackendTime
| extend IsError = ResponseCode >= 400

Ordering and Limiting

| order by TimeGenerated desc
| take 100
| top 10 by TotalTime desc

Common Scenarios

Daily Request Summary

ApiManagementGatewayLogs
| where TimeGenerated > ago(7d)
| summarize 
    TotalRequests = count(),
    SuccessfulRequests = countif(ResponseCode < 400),
    FailedRequests = countif(ResponseCode >= 400),
    AvgLatency = avg(TotalTime)
by bin(TimeGenerated, 1d), ApiId
| extend SuccessRate = round(100.0 * SuccessfulRequests / TotalRequests, 2)
| order by TimeGenerated desc

Top Consumers by Token Usage

let llmLogs = ApiManagementGatewayLlmLog 
| where TimeGenerated > ago(30d);

llmLogs 
| join kind=leftouter ApiManagementGatewayLogs on CorrelationId 
| summarize 
    TotalTokens = sum(TotalTokens),
    RequestCount = count()
by ApimSubscriptionId
| top 10 by TotalTokens desc

Error Rate Over Time

ApiManagementGatewayLogs
| where TimeGenerated > ago(24h)
| summarize 
    Total = count(),
    Errors = countif(ResponseCode >= 400)
by bin(TimeGenerated, 1h)
| extend ErrorRate = round(100.0 * Errors / Total, 2)
| project TimeGenerated, Total, Errors, ErrorRate
| order by TimeGenerated desc

Cache Hit Analysis

ApiManagementGatewayLogs
| where TimeGenerated > ago(24h)
| where Cache != ""
| summarize 
    HitCount = countif(Cache == "hit"),
    MissCount = countif(Cache == "miss")
by ApiId
| extend HitRate = round(100.0 * HitCount / (HitCount + MissCount), 2)

Streaming vs Non-Streaming LLM Requests

ApiManagementGatewayLlmLog
| where TimeGenerated > ago(24h)
| summarize 
    StreamingRequests = countif(IsStreamCompletion == true),
    NonStreamingRequests = countif(IsStreamCompletion == false),
    AvgTokens = avg(TotalTokens)
by DeploymentName

Running KQL Queries

Azure CLI

az monitor log-analytics query \
  -w <workspace-id> \
  --analytics-query "ApiManagementGatewayLogs | take 10"

Python

query = """
ApiManagementGatewayLogs
| where TimeGenerated > ago(1h)
| take 100
"""
result = utils.run(f'az monitor log-analytics query -w {workspace_id} --analytics-query "{query}"')

Best Practices

1. Always filter by time - Use where TimeGenerated > ago(...) to limit data scanned
2. Use summarize early - Aggregate data before joins when possible
3. Project only needed columns - Reduce data transfer with explicit projections
4. Use let statements - Break complex queries into readable parts
5. Join on CorrelationId - Link APIM tables together using this common key
6. Use bin() for time series - Group time data into meaningful intervals

References

• references/table-schemas.md - Complete table column definitions
• references/query-examples.md - Additional query examples from this repository

FinOps framework

The AI Gateway FinOps framework deploys two custom Log Analytics tables to enable cost tracking and budget enforcement per APIM subscription. Data is ingested via Data Collection Rules (DCR).

Use the query endpoint with API version 2020-08-01 to query the linked Log Analytics workspace. The results will be returned in PascalCase format.

PRICING_CL

Model pricing reference table. Stores per-model token prices that are periodically updated via DCR ingestion.

SUBSCRIPTION_QUOTA_CL

Per-subscription cost budget table. Stores the cost quota assigned to each APIM subscription.

Subscription budget vs actual cost

Joins LLM logs with pricing data and subscription quotas to calculate actual cost per subscription for the current month and compare it against the budget.

let llmHeaderLogs = ApiManagementGatewayLlmLog
| where TimeGenerated >= startofmonth(now()) and TimeGenerated <= endofmonth(now())
| where DeploymentName != "";
let llmLogsWithSubscriptionId = llmHeaderLogs
| join kind=leftouter ApiManagementGatewayLogs on CorrelationId
| project
    SubscriptionName = ApimSubscriptionId, DeploymentName, PromptTokens, CompletionTokens, TotalTokens;
llmLogsWithSubscriptionId
| join kind=inner (
    PRICING_CL
    | summarize arg_max(TimeGenerated, *) by Model
    | project Model, InputTokensPrice, OutputTokensPrice
    )
    on $left.DeploymentName == $right.Model
| extend InputCost = PromptTokens * InputTokensPrice
| extend OutputCost = CompletionTokens * OutputTokensPrice
| summarize
    InputCost = sum(InputCost), OutputCost = sum(OutputCost)
    by SubscriptionName
| extend TotalCost = (InputCost + OutputCost) / 1000
| join kind=inner (
    SUBSCRIPTION_QUOTA_CL
    | summarize arg_max(TimeGenerated, *) by Subscription
    | project Subscription, CostQuota
) on $left.SubscriptionName == $right.Subscription
| project SubscriptionName, CostQuota, TotalCost

This same query pattern is used by the FinOps framework's scheduled alert rules — one to suspend subscriptions that exceed their quota (TotalCost > CostQuota) and another to re-activate subscriptions that are back within budget (TotalCost <= CostQuota).

Source: Azure-Samples/ai-gateway-dev-portal — distributed by TomeVault.