WAP · Metrics Reference

Behind the health score
are these metrics

"Health Score: 98" is the output of dozens of metrics, continuously collected, converted, and weighted. This post lays out exactly what WAP tracks — from the host, to the instance, to the replication chain, to the query layer — one category at a time.

42%
CPU
68%
MEMORY
82%
DISK USAGE
43%
CONNECTIONS
91%
CACHE HIT RATE
6%
SLOW QUERY SHARE

Just 6 of them shown here · what WAP actually tracks goes far beyond this

The five categories below form the backbone of WAP's monitoring system — covering the full chain from hardware resources to application-level queries.

01 · Host-Level Metrics

The physical foundation the cluster runs on — any resource running dry shows up directly in database performance.

CPU Usage
Tracks user-space vs. system-space share to spot compute bottlenecks or runaway processes.
Memory Usage
Tracks physical memory and swap usage — low memory is usually the early warning sign for an OOM kill.
Disk IOPS / Throughput
Records read/write operation counts and bandwidth to pinpoint storage-layer bottlenecks.
Disk Usage
Continuously tracks remaining space to avoid an outage caused by a full disk.
Network Throughput
Monitors inbound and outbound NIC traffic to catch network-layer congestion or abnormal flows.

02 · Instance-Level Metrics

The MongoDB process's own runtime state — a direct read on the database's current load and health.

Connections
Active connections as a share of the maximum, warning early about connection exhaustion.
Opcounters
Per-second counts of insert / query / update / delete, mapping out the real shape of the workload.
Queue Length
How many reads and writes are waiting in queue — a sign the instance can't keep up.
Cache Hit Rate
WiredTiger cache hit ratio — a sudden drop usually means the working set has outgrown memory.
Page Eviction Rate
How often memory pages get evicted — an early signal of memory pressure.

03 · Replication & High Availability

Whether a replica set can fail over cleanly when something breaks depends entirely on this group of metrics staying healthy.

Replication Lag (Oplog Lag)
How far a secondary trails the primary — this directly sets the data-loss window a recovery can tolerate.
Election Frequency
Multiple elections in a short window are usually an early sign of network or resource trouble.
Heartbeat Latency
Round-trip time between node heartbeats — an abnormal rise can signal a network partition.
Replica Set Member State
Continuously tracks each member's health state and sync progress.

04 · Sharded Cluster Metrics

A set of metrics unique to sharded architecture — they determine whether load is actually spread evenly or not.

Shard Load Comparison
Request volume, CPU, and connections compared side by side across shards to catch hot shards early.
Chunk Distribution
Tracks how data chunks are distributed across shards, surfacing signs of imbalance.
Balancer Status
Whether the balancer is actively migrating chunks, or stuck.
Config Server Sync Lag
Health of the metadata replica set, which directly affects routing accuracy.

05 · Query & Performance Metrics

The metrics that ultimately decide whether users experience your database as "fast" or "slow."

Slow Query Share
The share of requests exceeding the latency threshold — a direct signal of performance degradation.
Query Latency Percentiles
P50 / P95 / P99 capture tail latency far more honestly than an average ever could.
Lock Wait Time
How long operations wait on lock contention — the core signal for concurrency bottlenecks.
Index Scan Efficiency
The ratio of documents scanned to documents returned — too low means the index isn't really earning its keep.
WAP Note

Collecting these metrics is only step one — WAP converts them into a readable health score, fires an alert the moment one crosses the safe line, and attaches the relevant context right alongside it, so an investigation never has to start from zero.

Ready?

Metrics aren't the point
being seen is

From the host to the query layer, Whaleal Platform puts the full metrics chain somewhere you'll actually look — instead of scattered across a dozen dashboards nobody opens.