Pipeline Bottleneck Analysis
Processing time + queue wait per pipeline step — identifies where documents spend the most timeTotal Avg Pipeline
17.0m
received → enrolled
Bottleneck Step
Patient Packet
avg 15.6m total
Bottleneck Processing
0s
avg actual work
Bottleneck Queue Wait
15.6m
100.0% of step time
Step Duration Breakdown — Avg Queue Wait + Avg Processing (seconds)
Per-Step Statistics
| Step | Statuses | Docs | — Processing Time (work) — | — Queue Wait (idle) — | Avg Total | Wait % | ||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Avg | Median | P95 | Max | Avg | Median | P95 | Max | |||||
| Image Extraction | 11→13 | 8 | 1s | 1s | 3s | 4s | 2s | 0s | 8s | 8s | 3s | 63.6% |
| Page Classification | 21→2x/83 | 0 | 0s | 0s | 0s | 0s | 0s | 0s | 0s | 0s | 0s | 0.0% |
| Entity Extraction | 31→3x | 0 | 0s | 0s | 0s | 0s | 0s | 0s | 0s | 0s | 0s | 0.0% |
| Patient Validation | 41→4x | 7 | 1.3m | 1.3m | 1.6m | 1.6m | 1s | 0s | 5s | 7s | 1.3m | 1.4% |
| BOTTLENECK Patient Packet | 51→5x | 7 | 0s | 0s | 1s | 1s | 15.6m | 0s | 1.3h | 1.8h | 15.6m | 100.0% |
| Enrollment Send | 61→6x | 6 | 0s | 0s | 0s | 0s | 0s | 0s | 0s | 0s | 0s | 33.3% |
Bottleneck Findings
Over 06/19/2026 – 06/19/2026 the average end-to-end pipeline time was 17.0m. The primary bottleneck is Patient Packet, averaging 15.6m per document (0s processing + 15.6m queue wait).
Queue Wait Dominates at Patient Packet (100.0% idle)
Documents wait 15.6m in queue before Patient Packet picks them up,
but actual processing only takes 0s.
The bottleneck is throughput capacity, not processing speed.
Increase Function App instances, reduce polling interval, or add Service Bus auto-scaling triggers.
High Queue Wait Ratio at Multiple Steps
51→5x
Patient Packet — 100.0% idle
(15.6m wait vs 0s work).
11→13 Image Extraction — 63.6% idle (2s wait vs 1s work).
Check Service Bus queue depths for these steps. Documents accumulating in queue means consumers are under-provisioned relative to producer output.
11→13 Image Extraction — 63.6% idle (2s wait vs 1s work).
Check Service Bus queue depths for these steps. Documents accumulating in queue means consumers are under-provisioned relative to producer output.
High P95/Median Ratio — Outlier Documents Present
Image Extraction — median 1s but P95 3s (2.8× median).
A small number of documents are significantly slower — likely large page counts, OCR retries, or external API timeouts. Consider a timeout/circuit-breaker at 3× the median to prevent these from blocking the pipeline.
A small number of documents are significantly slower — likely large page counts, OCR retries, or external API timeouts. Consider a timeout/circuit-breaker at 3× the median to prevent these from blocking the pipeline.