wu-sheng opened a new pull request, #13710:
URL: https://github.com/apache/skywalking/pull/13710
### Add QueueUsageBenchmark to validate BatchQueueStats metrics under
realistic backpressure
- [x] Tests(including UT, IT, E2E) are added to verify the new feature.
#### What
Add `QueueUsageBenchmark` — a separate benchmark that validates
`BatchQueueStats` usage metrics
under realistic backpressure by adding simulated per-item CPU cost
(busy-spin) in consumers.
Also add usage sampling to the existing `BatchQueueBenchmark`.
Unlike the existing `BatchQueueBenchmark` which uses no-op consumers (queue
stays near 0%),
this benchmark creates genuine backpressure so the queue fills up and the
usage percentage
becomes meaningful — validating the
`metrics_aggregation_queue_used_percentage` SO11Y metric.
#### Test scenarios
| Test | Consumer cost | Strategy | Types |
|------|--------------|----------|-------|
| `usageUnderLightLoad` | 200 ns/item | IF_POSSIBLE | 500 |
| `usageUnderMediumLoad` | 500 ns/item | IF_POSSIBLE | 500 |
| `usageUnderHeavyLoad` | 1 μs/item | IF_POSSIBLE | 500 |
| `usageUnderHeavyLoadBlocking` | 1 μs/item | BLOCKING | 500 |
| `usageUnderMediumLoad1000Types` | 500 ns/item | IF_POSSIBLE | 1000 |
| `usageUnderHeavyLoadBlocking1000Types` | 1 μs/item | BLOCKING | 1000 |
#### Benchmark results (8 threads, 32 producers, JDK 25)
```
=== Queue Usage Benchmark: light-load ===
Types: 500 Partitions: 350 BufferSize: 50000
Strategy: IF_POSSIBLE Consumer cost: 200 ns/item
Produced: 76,838,592 Consumed: 62,417,115
Consume rate: 12,483,423 items/sec
Drop rate: 18.77%
Total usage: min=45.6%, avg=83.1%, max=100.0%
Top partition: avg=100.0%, max=100.0%
=== Queue Usage Benchmark: medium-load ===
Types: 500 Partitions: 350 BufferSize: 50000
Strategy: IF_POSSIBLE Consumer cost: 500 ns/item
Produced: 58,351,493 Consumed: 38,301,507
Consume rate: 7,643,486 items/sec
Drop rate: 34.36%
Total usage: min=65.2%, avg=88.1%, max=100.0%
Top partition: avg=100.0%, max=100.0%
=== Queue Usage Benchmark: heavy-load ===
Types: 500 Partitions: 350 BufferSize: 50000
Strategy: IF_POSSIBLE Consumer cost: 1000 ns/item
Produced: 46,145,961 Consumed: 24,945,951
Consume rate: 4,989,190 items/sec
Drop rate: 45.94%
Total usage: min=42.1%, avg=91.4%, max=100.0%
Top partition: avg=98.7%, max=100.0%
=== Queue Usage Benchmark: heavy-load-blocking ===
Types: 500 Partitions: 350 BufferSize: 50000
Strategy: BLOCKING Consumer cost: 1000 ns/item
Produced: 38,724,600 Consumed: 31,502,928
Consume rate: 6,095,768 items/sec
Drop rate: 18.65%
Total usage: min=30.7%, avg=40.3%, max=52.8%
Top partition: avg=95.2%, max=100.0%
=== Queue Usage Benchmark: medium-load-1000t ===
Types: 1000 Partitions: 600 BufferSize: 50000
Strategy: IF_POSSIBLE Consumer cost: 500 ns/item
Produced: 72,106,253 Consumed: 38,796,897
Consume rate: 7,759,379 items/sec
Drop rate: 46.19%
Total usage: min=62.0%, avg=92.0%, max=100.0%
Top partition: avg=100.0%, max=100.0%
=== Queue Usage Benchmark: heavy-load-blocking-1000t ===
Types: 1000 Partitions: 600 BufferSize: 50000
Strategy: BLOCKING Consumer cost: 1000 ns/item
Produced: 51,913,900 Consumed: 32,526,678
Consume rate: 5,788,695 items/sec
Drop rate: 37.34%
Total usage: min=22.2%, avg=55.0%, max=67.1%
Top partition: avg=94.6%, max=100.0%
```
#### Key observations
- **Usage metrics validated**: `BatchQueueStats.totalUsedPercentage()`
correctly reports 40-92% avg under backpressure, confirming the SO11Y
`metrics_aggregation_queue_used_percentage` metric works as expected.
- **IF_POSSIBLE drops visible**: Drop rates scale with consumer cost (19% at
200ns → 46% at 1μs), confirming the strategy correctly sheds load.
- **BLOCKING backpressure**: Lower total usage (40-55% avg) because
producers block when partitions fill, throttling the produce rate.
- **Top partition saturation**: Individual partitions hit 100% even when
total usage is moderate, showing the value of per-partition monitoring via
`topN()`.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
