1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493# Equalify PDF Converter - Architecture Reference
## Overview
The Equalify PDF Converter is a monolithic Python application with background task queuing that transforms PDF course materials into accessible, semantic HTML for University of Illinois Chicago (UIC). The system enforces a strict architectural boundary: **course materials only** - no student records or personally identifiable information (PII) should reach processing stages.
**Architecture Pattern:** Monolith with Background Task Queue
- **Single Python Application:** One FastAPI process with multiple background worker threads
- **Asynchronous Processing:** Redis task queues enable non-blocking API responses
- **PII-First Design:** All documents undergo PII scanning before AI processing
- **Resilient Storage:** S3 operations use circuit breakers and exponential backoff retry
- **Containerized Everything:** Docker Compose for local development, AWS ECS for production
## Technology Stack
### Backend Framework
- **Python 3.11+** - Modern async/await support, type hints
- **FastAPI 0.115+** - High-performance async web framework
- **Uvicorn** - ASGI server with hot-reload support
- **Pydantic 2.0+** - Data validation and settings management
- **uv** - Fast Python package manager (replaces pip/poetry)
### AI/ML Processing
- **PydanticAI 1.0+** - Multi-agent AI orchestration framework
- **AWS Bedrock** - Claude AI models via AWS Bedrock
- **IBM Docling 2.55+** - PDF to Markdown conversion with OCR
- **Microsoft Presidio 2.2+** - PII detection and entity recognition
- **spaCy 3.7** - NLP models for Presidio (en_core_web_sm)
### Infrastructure
- **Redis 5.0+** - Task queues, job state, rate limiting, distributed locks
- **AWS S3** - Object storage for PDFs and results
- **LocalStack** - Local AWS emulation for development
- **Docker & Docker Compose** - Containerized services
### Monitoring & Observability
- **Prometheus** - Metrics collection and storage
- **Grafana** - Metrics visualization and dashboards
- **OpenTelemetry** - Distributed tracing instrumentation
- **prometheus-client** - Python metrics exporter
### Testing
- **pytest 7.4+** - Test framework
- **pytest-asyncio** - Async test support
- **pytest-xdist** - Parallel test execution
- **pytest-cov** - Coverage reporting
- **testcontainers** - Real Redis/S3 for integration tests
- **httpx** - Async HTTP client for API tests
- **reportlab** - PDF generation for test fixtures
### Deployment
- **AWS ECS Fargate** - Container orchestration (production)
- **AWS ElastiCache** - Managed Redis (production)
- **AWS CloudWatch** - Logging and monitoring (production)
- **Terraform** - Infrastructure as Code
## High-Level Architecture Diagram
```
┌─────────────────────────────────────────────────────────────┐
│ CLIENT APPLICATION │
│ (Web UI / API) │
└──────────────────────────┬──────────────────────────────────┘
│ HTTP/REST
▼
┌─────────────────────────────────────────────────────────────┐
│ FASTAPI APPLICATION │
│ ┌──────────────────────────────────────────────────────┐ │
│ │ API ENDPOINTS │ │
│ │ • POST /api/v1/documents/submit │ │
│ │ • GET /api/v1/documents/{id} │ │
│ │ • GET /api/v1/documents/{id}/result │ │
│ │ • GET /api/v1/approval/{token}/review │ │
│ │ • POST /api/v1/approval/{token}/decision │ │
│ │ • GET /health │ │
│ └──────────────────────────────────────────────────────┘ │
│ ┌──────────────────────────────────────────────────────┐ │
│ │ MIDDLEWARE STACK │ │
│ │ 1. CORS │ │
│ │ 2. Logging (request/response) │ │
│ │ 3. Rate Limiting (Redis sliding window) │ │
│ │ 4. Error Handler (global exception catching) │ │
│ │ 5. Metrics (Prometheus instrumentation) │ │
│ └──────────────────────────────────────────────────────┘ │
│ ┌──────────────────────────────────────────────────────┐ │
│ │ BACKGROUND WORKERS (Threads) │ │
│ │ • PII Worker (BLPOP eq-pdf:queue:pii) │ │
│ │ • Timeout Worker (Scheduled cleanup) │ │
│ │ • Processing runs via FastAPI BackgroundTasks │ │
│ └──────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────┘
│
│ Redis/S3/Bedrock
▼
┌─────────────────────────────────────────────────────────────┐
│ INFRASTRUCTURE SERVICES │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────────────┐ │
│ │ Redis │ │ AWS S3 │ │ AWS Bedrock │ │
│ │ │ │ │ │ (Claude AI) │ │
│ │ • Queues │ │ • Temp │ │ │ │
│ │ • Job State│ │ • Results │ │ • Haiku (fast) │ │
│ │ • Locks │ │ • Archive │ │ • Sonnet (quality) │ │
│ └─────────────┘ └─────────────┘ └─────────────────────┘ │
└─────────────────────────────────────────────────────────────┘
```
## Complete Directory Structure
```
src/
├── __init__.py
├── main.py # FastAPI app + lifespan (worker startup)
├── config.py # Pydantic Settings (env vars → typed config)
├── dependencies.py # FastAPI dependency injection providers
│
├── api/ # REST API Endpoints
│ ├── __init__.py
│ ├── documents.py # POST /submit, GET /status, GET /result
│ ├── approval.py # POST /approve, POST /reject
│ ├── health.py # GET /health, GET /metrics
│ └── dev_monitoring.py # Dev-only: GET /api/dev/monitoring/queues
│
├── workers/ # Background Task Processors
│ ├── __init__.py
│ ├── pii_worker.py # Monitors eq-pdf:queue:pii (Microsoft Presidio)
│ └── timeout_worker.py # Scheduled: approval timeouts, temp cleanup, orphan detection
│ # Note: Document processing runs via FastAPI BackgroundTasks, not a separate worker
│
├── services/ # Business Logic Layer
│ ├── __init__.py
│ ├── storage_service.py # S3 operations (upload/download/delete/presigned URLs)
│ ├── queue_service.py # Redis queue operations (LPUSH/BLPOP/LLEN)
│ ├── job_service.py # Job state management (Redis hashes + TTL)
│ ├── pii_service.py # PII detection orchestration
│ ├── pii_analyzer.py # Microsoft Presidio wrapper
│ ├── pipeline_viewer.py # Core 7-step versioned processing pipeline
│ ├── document_processing_service.py # Pipeline orchestration + S3/Redis integration
│ ├── pdf_extractor.py # Docling text extraction
│ ├── approval_service.py # Approval workflow (token validation, state transitions)
│ ├── timeout_service.py # Approval timeout monitoring (sorted sets)
│ ├── cleanup_service.py # Denied job file cleanup
│ ├── orphan_service.py # Orphaned job detection and cleanup
│ ├── s3_cleanup_service.py # S3 temporary file cleanup
│ ├── rate_limit_service.py # Rate limiting (Redis sliding window)
│ └── metrics_service.py # Prometheus metrics tracking
│
├── middleware/ # FastAPI Middleware
│ ├── __init__.py
│ ├── cors.py # CORS headers configuration
│ ├── logging_middleware.py # Request/response logging
│ ├── rate_limit.py # Rate limiting middleware (uses rate_limit_service)
│ ├── error_handler.py # Global exception handling and error formatting
│ └── metrics.py # Prometheus metrics collection
│
├── agents/ # AI Prompt Modules (PydanticAI)
│ ├── __init__.py
│ ├── model_tiers.py # Model selection (Haiku/Sonnet tier mapping)
│ └── prompts/ # Agent prompt templates (structure, boundary, footnote)
│
├── shared/ # Data Models and Constants
│ ├── __init__.py
│ ├── models/ # Pydantic Models
│ │ ├── __init__.py
│ │ ├── job.py # Job, JobCreate, JobUpdate, JobStatus
│ │ ├── queue.py # QueueMessage, PIIQueuePayload, ProcessingQueuePayload
│ │ ├── pii.py # PIIFinding, PIIResult, PIIApprovalRequest
│ │ ├── processing.py # ProcessingResult, ProcessingError
│ │ ├── approval.py # ApprovalRequest, ApprovalResponse
│ │ └── redis_schema.py # Redis data structure schemas
│ └── constants/ # Application Constants
│ ├── __init__.py
│ ├── queues.py # Queue names (QUEUE_PII, QUEUE_PROCESSING)
│ ├── redis_keys.py # Redis key patterns (job:*, timeout:*, metrics:*)
│ └── statuses.py # Job status constants (SUBMITTED, PII_SCANNING, etc.)
│
└── utils/ # Utility Functions
├── __init__.py
├── circuit_breaker.py # Circuit breaker pattern for S3 resilience
├── confidence_scoring.py # AI quality confidence calculation
├── text_cleanup.py # Text normalization and cleanup
├── token_generator.py # Secure approval token generation
└── retry_helpers.py # Exponential backoff retry decorators
```
## Application Layers
### 1. API Layer (FastAPI)
**Technology:** FastAPI 0.115+
**Purpose:** REST API endpoints for document submission, status tracking, results retrieval, and approval workflow
**Pattern:** Async request handlers with dependency injection
#### Files
**`api/documents.py`** - Document lifecycle management
- `POST /api/v1/documents/submit` - Submit PDF for processing (returns job_id)
- `GET /api/v1/documents/{job_id}` - Check processing status
- `GET /api/v1/documents/{job_id}/result` - Retrieve processed HTML/MDX with presigned URLs
**`api/approval.py`** - Human approval workflow
- `GET /api/v1/approval/{token}/review` - Get job details for PII review (requires token)
- `POST /api/v1/approval/{token}/decision` - Submit approval/denial decision (requires token)
**`api/health.py`** - Health checks and metrics
- `GET /health` - Application health status (Redis/S3 connectivity)
- `GET /metrics` - Prometheus metrics endpoint
**`api/dev_monitoring.py`** - Development-only endpoints (disabled in production)
- `GET /api/dev/monitoring/queues` - Real-time queue depths and job counts
### 2. Service Layer (Business Logic)
**Technology:** Pure Python classes with async/await
**Purpose:** Shared business logic, external service integration, data orchestration
**Pattern:** Dependency injection via FastAPI `Depends()`
#### Storage & State Management
**`services/storage_service.py`** - S3 operations
- Upload PDFs to temp bucket, results to permanent bucket
- Download files for processing
- Generate presigned URLs for client downloads
- Circuit breakers for resilience (upload/download/delete circuits)
- Exponential backoff retry on transient errors
**`services/queue_service.py`** - Redis queue operations
- `enqueue()` - LPUSH to Redis lists (eq-pdf:queue:pii, eq-pdf:queue:processing)
- `dequeue()` - BLPOP with timeout for graceful shutdown
- `queue_depth()` - LLEN for monitoring
**`services/job_service.py`** - Job state management
- Create job metadata in Redis hashes (eq-pdf:job:{job_id})
- Update status with automatic TTL (7 days active, 30 days completed)
- Retrieve job details with status filtering
- Atomic state transitions
#### PII Detection
**`services/pii_service.py`** - PII detection orchestration
- Coordinate PDF download, text extraction, Presidio scanning
- Handle approval workflow initiation
- Store PII findings in job metadata
**`services/pii_analyzer.py`** - Microsoft Presidio wrapper
- Initialize Presidio AnalyzerEngine with spaCy models
- Detect entities (PERSON, EMAIL_ADDRESS, PHONE_NUMBER, SSN, etc.)
- Return structured PII findings with confidence scores
#### Document Processing
**`services/processing_service.py`** - Main processing pipeline
- Orchestrate PDF → Markdown → AI-enhanced HTML workflow
- Manage multi-stage processing with error recovery
- Calculate confidence scores
- Store results in S3 with versioning
**`services/pdf_converter.py`** - Docling PDF → Markdown
- Convert PDF to Markdown with page images
- Handle multi-page documents
- OCR for scanned documents
- Preserve document structure (headings, lists, tables)
**`services/pdf_extractor.py`** - Docling text extraction
- Extract raw text for PII scanning (faster than full conversion)
- Normalize whitespace and formatting
**`services/ai_enhancement_service.py`** - AI accessibility enhancement
- Process pages concurrently (configurable concurrency)
- Add alt text to images
- Improve heading hierarchy
- Add ARIA labels for semantic structure
- Retry failed pages with exponential backoff
#### Approval & Cleanup
**`services/approval_service.py`** - Approval workflow
- Generate secure approval tokens (cryptographically random)
- Validate tokens and approval state
- Transition jobs from awaiting_approval → processing (approved) or denied (rejected)
- Trigger cleanup for denied jobs
**`services/timeout_service.py`** - Approval timeout monitoring
- Monitor Redis sorted set for expired approvals (eq-pdf:timeouts:approval)
- Auto-reject documents after timeout (default: 4 hours)
- Remove from timeout queue on approval/rejection
**`services/cleanup_service.py`** - Denied job cleanup
- Delete temp files from S3 when job is rejected
- Remove job metadata from Redis
- Log cleanup actions for audit trail
**`services/s3_cleanup_service.py`** - S3 temporary file cleanup
- Periodic cleanup of temp files older than retention period (default: 24 hours)
- List and delete in batches for efficiency
- Track cleanup metrics
**`services/orphan_service.py`** - Orphaned job detection
- Detect jobs stuck in processing state beyond max processing time (default: 2 hours)
- Mark as failed with timeout reason
- Clean up orphaned S3 files
#### Infrastructure Services
**`services/rate_limit_service.py`** - Rate limiting
- Redis sliding window algorithm
- Per-IP and global rate limits
- Configurable time windows and request limits
**`services/metrics_service.py`** - Prometheus metrics
- Track job counts by status
- Monitor queue depths
- Record processing times
- S3 operation metrics (success/failure/circuit_breaker state)
- Cleanup operation metrics
### 3. Worker Layer (Background Processing)
**Technology:** Asyncio tasks managed by FastAPI lifespan
**Purpose:** Asynchronous background job processing
**Pattern:** Blocking queue consumers with graceful shutdown
#### Files
**`workers/pii_worker.py`** - PII detection worker
- `BLPOP eq-pdf:queue:pii` - Wait for PII scan jobs
- Download PDF from S3 temp bucket
- Run Microsoft Presidio scan
- **If clean:** Queue for processing (LPUSH eq-pdf:queue:processing)
- **If PII found:** Update status to awaiting_approval, add to timeout queue
- Graceful shutdown on signal (completes current job, max 30s)
**Document Processing** (via FastAPI BackgroundTasks)
- `DocumentProcessingService.process_document()` runs as background task
- Downloads PDF from S3
- `PipelineViewerService` runs 7-step versioned pipeline (Docling → Structure → Headings → Pages → Code Blocks → Boundaries → Cleanup)
- Stores final markdown + figures to S3 results bucket
- Updates job state to "completed" with cost/token metadata
**`workers/timeout_worker.py`** - Scheduled maintenance worker
- **Approval timeout checks** (every 30 seconds)
- `ZRANGEBYSCORE eq-pdf:timeouts:approval` - Find expired approvals
- Auto-reject and cleanup
- **Temp file cleanup** (every 1 hour)
- Delete S3 temp files older than retention period
- **Orphan detection** (every 4 hours)
- Find jobs stuck in processing state
- Mark as failed, cleanup S3 files
- **Metrics cleanup** (every 24 hours)
- Aggregate and archive old metrics
- Graceful shutdown on signal
### 4. Middleware Layer (Cross-Cutting Concerns)
**Technology:** FastAPI middleware stack
**Purpose:** Request/response interception for logging, security, metrics
**Pattern:** Starlette BaseHTTPMiddleware
#### Execution Order
FastAPI middleware executes in reverse order of registration (last added = first executed):
1. **`middleware/cors.py`** - CORS headers
- Allow origins, methods, headers for web clients
- Configurable for dev (localhost) vs production domains
2. **`middleware/logging_middleware.py`** - Request/response logging
- Log request method, path, status, duration
- Structured JSON logging for CloudWatch
3. **`middleware/rate_limit.py`** - Rate limiting enforcement
- Check rate limits before processing request
- Return 429 Too Many Requests if exceeded
- Uses `RateLimitService` (Redis sliding window)
4. **`middleware/error_handler.py`** - Global exception handling
- Catch all unhandled exceptions
- Return structured error responses (RFC 7807 Problem Details)
- Log errors with stack traces
- Hide internal errors from clients (security)
5. **`middleware/metrics.py`** - Prometheus instrumentation
- Track request counts by method/endpoint/status
- Record request duration histograms
- Monitor active requests gauge
### 5. Data Models Layer (Pydantic)
**Technology:** Pydantic 2.0+
**Purpose:** Type-safe data validation, serialization, API contracts
**Pattern:** BaseModel with Field validators
#### Files
**`shared/models/job.py`** - Job lifecycle models
- `Job` - Complete job with all fields (status, s3_key, result URLs, timestamps)
- `JobCreate` - Minimal fields for job creation (filename, content_type)
- `JobUpdate` - Partial update model (status, s3_key, error_message)
- `JobStatus` - Read-only status response for clients
**`shared/models/queue.py`** - Queue message payloads
- `QueueMessage` - Base message (job_id, timestamp)
- `PIIQueuePayload` - PII scan job (job_id, s3_key)
- `ProcessingQueuePayload` - Processing job (job_id, s3_key, approved_at)
**`shared/models/pii.py`** - PII detection models
- `PIIFinding` - Presidio entity result (entity_type, start, end, score)
- `PIIResult` - Scan result (has_pii, findings, confidence)
- `PIIApprovalRequest` - Approval action (job_id, decision, justification)
**`shared/models/processing.py`** - Processing results
- `ProcessingResult` - Complete result (markdown_url, confidence_score)
- `ProcessingError` - Structured error (stage, error_type, message, retryable)
**`shared/models/approval.py`** - Approval workflow
- `ApprovalRequest` - Client request (decision, justification, token)
- `ApprovalResponse` - Server response (accepted, new_status, message)
**`shared/models/redis_schema.py`** - Redis data structure documentation
- Type definitions for Redis hash fields
- Key patterns and naming conventions
- TTL policies
### 6. Constants Layer (Application Configuration)
**Technology:** Python constants
**Purpose:** Centralized constant definitions, prevent magic strings
**Pattern:** Module-level constants
#### Files
**`shared/constants/statuses.py`** - Job status constants
```python
SUBMITTED = "submitted"
PII_SCANNING = "pii_scanning"
AWAITING_APPROVAL = "awaiting_approval"
PROCESSING = "processing"
COMPLETED = "completed"
FAILED = "failed"
DENIED = "denied"
```
**`shared/constants/queues.py`** - Queue names
```python
QUEUE_PII = "eq-pdf:queue:pii"
QUEUE_PROCESSING = "eq-pdf:queue:processing"
```
**`shared/constants/redis_keys.py`** - Redis key patterns
```python
JOB_PREFIX = "eq-pdf:job:"
TIMEOUT_PREFIX = "eq-pdf:timeouts:"
METRICS_PREFIX = "eq-pdf:metrics:"
LOCK_PREFIX = "eq-pdf:lock:"
```
### 7. Utilities Layer (Helper Functions)
**Technology:** Pure Python utility functions
**Purpose:** Reusable logic for retry, scoring, text processing
**Pattern:** Stateless functions
#### Files
**`utils/circuit_breaker.py`** - Circuit breaker pattern
- `CircuitBreaker` class with states (CLOSED, OPEN, HALF_OPEN)
- Track consecutive failures (threshold: 5)
- Timeout before recovery test (60 seconds)
- Separate circuits for S3 upload/download/delete
**`utils/retry_helpers.py`** - Exponential backoff retry
- `retry_with_backoff_for_sync_func()` - Async wrapper for synchronous functions with exponential backoff
- `retry_with_backoff()` - Async version for async functions
- Categorize errors as retriable vs non-retriable
- Max attempts: 3, base delay: 1s, max delay: 8s
**`utils/confidence_scoring.py`** - AI quality scoring
- Calculate overall confidence from page-level scores
- Weight by page length
- Classify as HIGH (>0.85), MEDIUM (0.60-0.85), LOW (<0.60)
**`utils/text_cleanup.py`** - Text normalization
- Remove excessive whitespace
- Normalize line endings
- Fix common PDF extraction artifacts
- Preserve code blocks and preformatted text
**`utils/token_generator.py`** - Secure token generation
- Generate cryptographically secure random tokens for approval URLs
- URL-safe base64 encoding
- Configurable token length (default: 32 bytes → 43 characters)
## Processing Pipeline: Stage-by-Stage Flow
### Stage 1: Document Submission
**Endpoint:** `POST /api/v1/documents/submit`
**Duration:** <100ms
**Status:** `submitted` → `pii_scanning`
#### Flow
1. Client uploads PDF file (max 100MB)
2. API validates:
- Content type is `application/pdf`
- File size within limit
- Basic PDF header check
3. Generate job_id (UUID4)
4. Upload PDF to S3 temp bucket: `s3://equalify-pdf-temp/{job_id}.pdf`
5. Create job in Redis:
```python
HSET eq-pdf:job:{job_id}
status "pii_scanning"
s3_key "temp/{job_id}.pdf"
filename "course-syllabus.pdf"
created_at "2025-01-15T10:00:00Z"
```
6. Queue for PII scanning:
```python
LPUSH eq-pdf:queue:pii '{"job_id": "...", "s3_key": "temp/..."}'
```
7. Return immediately:
```json
{
"job_id": "abc-123-def-456",
"status": "pii_scanning",
"estimated_completion_minutes": 5
}
```
#### Edge Cases
- **File too large (>100MB):** Return `413 Payload Too Large`
- **Invalid PDF:** Return `400 Bad Request` with validation error
- **S3 upload failure:** Retry 3 times, then return `503 Service Unavailable`
- **Redis unavailable:** Return `503 Service Unavailable` (can't track job)
### Stage 2: PII Detection & Scanning
**Worker:** `pii_worker.py`
**Duration:** 10-30 seconds
**Status:** `pii_scanning` → `processing` OR `awaiting_approval`
#### Flow
1. PII worker blocks on queue:
```python
job = BLPOP eq-pdf:queue:pii 60 # 60s timeout for graceful shutdown
```
2. Download PDF from S3 temp bucket
3. Extract text using Docling (fast extraction, no images)
4. Run Microsoft Presidio scan:
- Detect entities: PERSON, EMAIL_ADDRESS, PHONE_NUMBER, SSN, CREDIT_CARD, etc.
- Return findings with confidence scores (0.0-1.0)
5. **Decision point:**
**If NO PII detected** (findings empty):
- Update job status to `processing`
- Queue for processing:
```python
LPUSH eq-pdf:queue:processing '{"job_id": "...", "s3_key": "..."}'
```
**If PII detected** (findings exist):
- Generate secure approval token
- Update job:
```python
HSET eq-pdf:job:{job_id}
status "awaiting_approval"
pii_findings '[{"entity_type": "PERSON", "score": 0.95, ...}]'
approval_token "secure-random-token"
expires_at "2025-01-15T14:00:00Z" # +4 hours
```
- Add to timeout queue:
```python
ZADD eq-pdf:timeouts:approval 1705329600 "{job_id}" # Unix timestamp
```
- Send email notification to faculty (future phase)
#### Edge Cases
- **PII scan timeout (>60s):** Retry once, then mark as failed
- **Presidio service unavailable:** Queue for retry with exponential backoff
- **False positives:** Common patterns (e.g., "CS 101") added to allowlist
- **Text extraction failure:** Mark as failed with error message
- **S3 download failure:** Retry with exponential backoff, then fail
### Stage 3: Human Approval Workflow
**Endpoint:** `POST /api/v1/documents/{job_id}/approve` OR `/reject`
**Duration:** Variable (human decision)
**Status:** `awaiting_approval` → `processing` (approved) OR `denied` (rejected)
#### Approval Flow
1. Faculty receives notification with:
- Job ID
- PII findings (entity types, redacted snippets)
- Secure approval URL with token
2. Faculty reviews findings and decides:
- **Approve:** Course material, false positive, acceptable risk
- **Reject:** Contains PII, not appropriate for processing
3. API validates:
- Job exists and status is `awaiting_approval`
- Approval token matches stored token
- Not expired (within 4-hour window)
4. **If approved:**
- Update status to `processing`
- Remove from timeout queue: `ZREM eq-pdf:timeouts:approval {job_id}`
- Queue for processing: `LPUSH eq-pdf:queue:processing {...}`
- Log approval decision with justification
5. **If rejected:**
- Update status to `denied`
- Remove from timeout queue
- Trigger cleanup: Delete S3 temp file, remove job metadata after 7 days
- Log rejection decision with justification
#### Edge Cases
- **Approval timeout (>4 hours):** Timeout worker auto-rejects and cleans up
- **Invalid token:** Return `403 Forbidden`
- **Double approval:** First valid approval wins, subsequent ignored
- **Already processed:** Return `409 Conflict` (status changed)
- **Email delivery failure:** Log warning, user can check status via API
### Stage 4: Document Processing
**Service:** `DocumentProcessingService` + `PipelineViewerService`
**Trigger:** FastAPI BackgroundTasks (inline, not a separate worker)
**Duration:** 2-8 minutes
**Status:** `processing` → `completed` OR `failed`
#### Flow
1. `DocumentProcessingService.process_document()` runs as a background task
2. Download PDF from S3 temp bucket
3. **PipelineViewerService** runs 7-step versioned pipeline:
- **Step 1 - Docling extraction (v0):** PDF → markdown + page images (30-90s)
- **Step 2 - Structure analysis:** AI identifies headings, footnotes, page types per-page
- **Step 3 - Heading level fix:** Deterministic heading hierarchy normalization
- **Step 4 - Page content corrections (v1):** AI fixes OCR errors per-page (parallel, semaphore-limited)
- **Step 5 - Code block tagging:** Identify programming languages in code blocks
- **Step 6 - Cross-page boundary fixes (v2):** Rejoin split content, relocate footnotes
- **Step 7 - Final cleanup (v3):** Normalize whitespace and formatting
4. Each step produces a versioned snapshot (v0, v1, v2, v3) with complete markdown for review
5. **Store results in S3:**
- Upload final markdown + extracted figures to results bucket
- Store change ledger (audit trail of all AI edits)
6. **Update job:**
```python
HSET eq-pdf:job:{job_id}
status "completed"
markdown_url "jobs/{job_id}/final.md"
confidence_score "0.87"
llm_cost_cents "5.25"
completed_at "2025-01-15T10:05:23Z"
```
#### Edge Cases
- **Processing timeout (>10 minutes):** Kill job, mark as failed
- **Individual page AI failures:** Log error, keep original markdown for that page
- **S3 upload failure:** Retry 3 times, then mark job as failed
- **Malformed PDF:** Docling failures handled, return structured error
- **Circuit breaker open (S3 degraded):** Fail fast, retry later
### Stage 5: Result Delivery
**Endpoint:** `GET /api/v1/documents/{job_id}/result`
**Duration:** <50ms
**Status:** `completed` (no change)
#### Flow
1. Client polls for result (or uses webhook in future phase)
2. API retrieves job from Redis: `HGET eq-pdf:job:{job_id}`
3. Validate status is `completed`
4. Return result:
```json
{
"job_id": "abc-123-def-456",
"status": "completed",
"markdown_url": "https://s3.../result.md?X-Amz-...",
"confidence_score": 0.87,
"review_recommended": false,
"processing_time_seconds": 243,
"completed_at": "2025-01-15T10:05:23Z"
}
```
5. Client downloads files from presigned URLs
6. Result URLs expire after 7 days (regenerate if needed)
#### Edge Cases
- **Job not found:** Return `404 Not Found`
- **Job still processing:** Return `202 Accepted` with status and ETA
- **Job failed:** Return `200 OK` with status `failed` and error message
- **Presigned URL expired:** Regenerate on demand (requires S3 access)
## Background Services: Scheduled Maintenance
### Timeout Cleanup Worker
**Schedule:** Every 30 seconds
**Purpose:** Enforce approval deadlines, prevent indefinite holds
#### Operations
1. **Check expired approvals:**
```python
expired_jobs = ZRANGEBYSCORE eq-pdf:timeouts:approval 0 {current_timestamp}
```
2. For each expired job:
- Check job status (race condition protection)
- If still `awaiting_approval`:
- Update status to `denied`
- Delete S3 temp file
- Remove from timeout queue
- Log auto-rejection event
3. **Check stuck processing jobs:**
- Find jobs in `processing` state older than 2 hours
- Mark as `failed` with timeout reason
- Cleanup S3 files
#### Edge Cases
- **Redis unavailable:** Skip cycle, log warning, retry next cycle
- **S3 deletion failure:** Log error, mark for manual cleanup, continue
- **Race condition (job approved during cleanup):** Check status again before deletion
### Temporary File Cleanup
**Schedule:** Every 1 hour
**Purpose:** Prevent S3 storage bloat from temp files
#### Operations
1. List files in temp bucket older than retention period (24 hours)
2. Filter files not associated with active jobs:
- Extract job_id from S3 key
- Check job status in Redis
- Skip if job is `pii_scanning`, `awaiting_approval`, or `processing`
3. Delete in batches (max 1000 per batch)
4. Track metrics: files deleted, bytes freed, errors
### Orphan Detection
**Schedule:** Every 4 hours
**Purpose:** Detect and cleanup orphaned jobs and data
#### Operations
1. **Find orphaned jobs:**
- Jobs in `processing` state older than `max_processing_hours` (2 hours)
- Jobs with S3 files but no Redis metadata
- Redis jobs with no corresponding S3 files
2. **Cleanup actions:**
- Mark stuck jobs as `failed`
- Delete orphaned S3 files
- Remove orphaned Redis keys
3. **Alert on high orphan rate** (>5% of jobs)
## State Machine: Job Status Transitions
```mermaid
stateDiagram-v2
[*] --> submitted: POST /submit
submitted --> pii_scanning: Queue for PII
pii_scanning --> processing: No PII detected
pii_scanning --> awaiting_approval: PII detected
pii_scanning --> failed: Scan error
awaiting_approval --> processing: Approved
awaiting_approval --> denied: Rejected
awaiting_approval --> denied: Timeout
processing --> completed: Success
processing --> failed: Error/Timeout
completed --> [*]
failed --> [*]
denied --> [*]
```
### Status Definitions
- **`submitted`** - Initial state, PDF uploaded to S3, queued for PII
- **`pii_scanning`** - PII worker processing, Microsoft Presidio scan in progress
- **`awaiting_approval`** - PII detected, waiting for human decision (4-hour timeout)
- **`processing`** - Approved or clean, Docling + AI enhancement in progress
- **`completed`** - Success, HTML/MDX stored in S3, results available
- **`failed`** - Error occurred (scan failure, processing timeout, S3 error)
- **`denied`** - Rejected by faculty or auto-rejected on timeout
### Terminal States
- **`completed`** - TTL: 30 days in Redis, permanent in S3
- **`failed`** - TTL: 30 days in Redis, no S3 results
- **`denied`** - TTL: 7 days in Redis, S3 temp files deleted immediately
## Redis Data Structures: Deep Dive
### Application Namespace: `eq-pdf:*`
All Redis keys use the `eq-pdf:` prefix for namespacing.
### 1. Task Queues (Redis Lists)
Used for background job processing with blocking operations.
```python
# PII scanning queue
eq-pdf:queue:pii = [
'{"job_id": "abc-123", "s3_key": "temp/abc-123.pdf", "created_at": "2025-01-15T10:00:00Z"}',
'{"job_id": "def-456", "s3_key": "temp/def-456.pdf", "created_at": "2025-01-15T10:01:00Z"}'
]
# Processing queue (approved or clean documents)
eq-pdf:queue:processing = [
'{"job_id": "ghi-789", "s3_key": "temp/ghi-789.pdf", "approved_at": "2025-01-15T10:05:00Z"}',
'{"job_id": "jkl-012", "s3_key": "temp/jkl-012.pdf", "approved_at": null}'
]
```
**Operations:**
- **Enqueue:** `LPUSH eq-pdf:queue:pii '{...}'` (left push, FIFO order)
- **Dequeue:** `BRPOP eq-pdf:queue:pii 60` (blocking right pop, 60s timeout)
- **Depth:** `LLEN eq-pdf:queue:pii` (queue length for monitoring)
**Why Lists?**
- FIFO ordering (documents processed in submission order)
- Blocking operations (workers sleep when idle, wake on new jobs)
- Atomic operations (no race conditions)
- Persistence (jobs not lost on restart)
### 2. Job State (Redis Hashes)
Each job stores metadata in a Redis hash with automatic TTL.
```python
# Active job (pii_scanning, awaiting_approval, processing)
eq-pdf:job:abc-123 = {
"status": "awaiting_approval",
"s3_key": "temp/abc-123.pdf",
"filename": "course-syllabus.pdf",
"content_type": "application/pdf",
"created_at": "2025-01-15T10:00:00Z",
"updated_at": "2025-01-15T10:00:45Z",
"pii_findings": '[{"entity_type": "PERSON", "score": 0.95}]',
"approval_token": "secure-random-token-here",
"expires_at": "2025-01-15T14:00:45Z"
}
# Completed job
eq-pdf:job:def-456 = {
"status": "completed",
"s3_key": "temp/def-456.pdf",
"filename": "lecture-notes.pdf",
"markdown_url": "https://s3.../result.md?X-Amz-...",
"confidence_score": "0.87",
"created_at": "2025-01-15T10:01:00Z",
"completed_at": "2025-01-15T10:05:23Z"
}
# Failed job
eq-pdf:job:ghi-789 = {
"status": "failed",
"s3_key": "temp/ghi-789.pdf",
"error_message": "PDF extraction failed: Encrypted document",
"error_stage": "pdf_conversion",
"created_at": "2025-01-15T10:02:00Z",
"failed_at": "2025-01-15T10:02:15Z"
}
```
**TTL Policy:**
- **Active jobs** (`pii_scanning`, `awaiting_approval`, `processing`): 7 days
- **Completed jobs:** 30 days
- **Failed jobs:** 30 days
- **Denied jobs:** 7 days
**Operations:**
- **Create:** `HSET eq-pdf:job:{job_id} status submitted ... | EXPIRE eq-pdf:job:{job_id} 604800`
- **Update:** `HSET eq-pdf:job:{job_id} status processing updated_at ...`
- **Read:** `HGETALL eq-pdf:job:{job_id}`
- **Delete:** `DEL eq-pdf:job:{job_id}` (or let TTL expire)
**Why Hashes?**
- Efficient storage (fields stored inline, not separate keys)
- Atomic field updates (status transitions without race conditions)
- Flexible schema (add fields as needed, no migration)
- TTL support (automatic cleanup, prevent memory leaks)
### 3. Timeout Tracking (Redis Sorted Sets)
Track approval deadlines with automatic expiration.
```python
# Sorted set: score = Unix timestamp (expiration time)
eq-pdf:timeouts:approval = {
"abc-123": 1705329645.0, # Expires at 2025-01-15 14:00:45 UTC
"def-456": 1705330200.0, # Expires at 2025-01-15 14:10:00 UTC
"ghi-789": 1705331000.0 # Expires at 2025-01-15 14:23:20 UTC
}
```
**Operations:**
- **Add deadline:** `ZADD eq-pdf:timeouts:approval {timestamp} {job_id}`
- **Find expired:** `ZRANGEBYSCORE eq-pdf:timeouts:approval 0 {current_time}`
- **Remove:** `ZREM eq-pdf:timeouts:approval {job_id}`
- **Count pending:** `ZCARD eq-pdf:timeouts:approval`
**Why Sorted Sets?**
- Efficient range queries (find expired in O(log N + M))
- Automatic ordering (no need to sort)
- Score-based expiration (Unix timestamps as scores)
- Atomic operations (add/remove without locks)
### 4. Distributed Locks (Redis Strings with NX and EX)
Prevent duplicate processing and race conditions.
```python
# Lock format: eq-pdf:lock:{resource}
eq-pdf:lock:job:abc-123 = "worker-instance-id" # EX 60 (expires in 60 seconds)
```
**Operations:**
- **Acquire:** `SET eq-pdf:lock:job:{job_id} {worker_id} NX EX 60`
- **Release:** `DEL eq-pdf:lock:job:{job_id}` (only if owner)
**Use Cases:**
- Prevent duplicate PII scans (if job re-queued)
- Prevent duplicate processing (if approval timeout and manual approval race)
- Cleanup operations (ensure one worker cleans temp files)
### 5. Metrics (Redis Hashes)
Track daily/weekly metrics for monitoring and billing.
```python
# Daily metrics (key: eq-pdf:metrics:daily:2025-01-15)
eq-pdf:metrics:daily:2025-01-15 = {
"submissions": 45,
"completed": 38,
"failed": 2,
"denied": 5,
"pii_detected": 12,
"pii_clean": 33,
"avg_processing_time_seconds": 243,
"total_pages_processed": 1234
}
# Weekly metrics (aggregated from daily)
eq-pdf:metrics:weekly:2025-W03 = {
"submissions": 312,
"completed": 285,
...
}
```
**TTL Policy:**
- Daily metrics: 90 days
- Weekly metrics: 2 years
## S3 Bucket Organization
### Temp Bucket: `equalify-pdf-temp`
**Purpose:** Store unvalidated PDFs during PII scanning and approval
**Lifecycle:** Files deleted after 24 hours OR when job completes/denied
**Access:** Private (API and workers only)
```
equalify-pdf-temp/
├── temp/
│ ├── abc-123-def-456.pdf # Original uploaded PDF
│ ├── ghi-789-jkl-012.pdf
│ └── ...
```
**Lifecycle Policy:**
- Delete files older than 24 hours (backup cleanup)
- No versioning (temp files, not critical)
### Results Bucket: `equalify-pdf-results`
**Purpose:** Store processed HTML/MDX results
**Lifecycle:** Permanent storage with versioning
**Access:** Public read (presigned URLs for authenticated users)
```
equalify-pdf-results/
├── abc-123-def-456/
│ ├── final.md # Final semantic markdown
│ ├── ledger.json # Change ledger (audit trail of AI edits)
│ └── metadata.json # Processing metadata (confidence, timestamps)
├── ghi-789-jkl-012/
│ ├── result.html
│ ├── result.mdx
│ └── metadata.json
└── ...
```
**Lifecycle Policy:**
- Transition to Glacier after 1 year (cost optimization)
- Versioning enabled (support document updates)
- Server-side encryption (AES-256)
## Infrastructure: Local vs Production
### Comparison Table
| Component | Local Development | Production (AWS) |
|-----------|------------------|------------------|
| **Redis** | Docker container (`redis:latest`) | AWS ElastiCache (Redis 7.x, cluster mode) |
| **S3** | LocalStack container (port 4566) | AWS S3 (multi-region replication) |
| **AI Provider** | AWS Bedrock (us-east-1) OR Anthropic API | AWS Bedrock (us-east-1, production account) |
| **Monitoring** | Prometheus + Grafana (Docker) | AWS CloudWatch Logs + Metrics + Alarms |
| **Orchestration** | Docker Compose | AWS ECS Fargate (spot instances for cost) |
| **Networking** | Docker bridge network | AWS VPC with private subnets |
| **Load Balancer** | None (direct container access) | AWS ALB with health checks |
| **Secrets** | `.env` file (NOT committed) | AWS Secrets Manager |
| **IAM** | Static credentials (test/test) | ECS Task Role (no credentials in code) |
| **Logs** | Container stdout/stderr | CloudWatch Logs with structured JSON |
| **Domains** | localhost:8080 | equalify-pdf.uic.edu (Route 53) |
| **SSL/TLS** | None (local HTTP) | ACM certificate + ALB HTTPS termination |
### Local Development Setup
**Prerequisites:**
- Docker Desktop
- 8GB+ RAM (for LocalStack + containers)
- 20GB+ disk space
**Start services:**
```bash
make dev
# Equivalent to:
# docker-compose -f docker-compose.yml -f docker-compose.dev.yml up -d
```
**Services started:**
- **api-gateway** (port 8080) - FastAPI application
- **redis** (port 6379) - Task queues and state
- **localstack** (port 4566) - Local AWS emulation
- **demo-ui** (port 5173) - React demo frontend
- **prometheus** (port 9090) - Metrics collection
- **grafana** (port 3001) - Metrics visualization
- **redis-exporter** (port 9121) - Redis metrics
**Access points:**
- API: <http://localhost:8080>
- API Docs: <http://localhost:8080/docs>
- Metrics: <http://localhost:8080/metrics>
- Demo UI: <http://localhost:5173>
- Grafana: <http://localhost:3001> (admin/admin)
- Prometheus: <http://localhost:9090>
**Hot reload:**
- Source code mounted as read-only volume
- Uvicorn `--reload` watches for changes
- Tests mounted for containerized test runs
**Detailed setup:** See [docs/environment-setup-simplified.md](environment-setup-simplified.md)
### Production Deployment (AWS)
**Architecture:**
- **ECS Fargate:** Single task definition (API + workers), spot instances
- **ElastiCache:** Redis cluster (multi-AZ, automatic failover)
- **S3:** Two buckets (temp, results) with lifecycle policies
- **ALB:** Application Load Balancer with health checks
- **CloudWatch:** Logs, metrics, alarms
- **Secrets Manager:** API keys, database credentials
**Deployment process:**
1. Build Docker image with production target
2. Push to AWS ECR (Elastic Container Registry)
3. Update ECS task definition with new image tag
4. ECS performs rolling update (zero-downtime)
5. Health checks validate new tasks before switching traffic
**Infrastructure as Code:** See [scripts/deploy-aws.sh](../scripts/deploy-aws.sh) and [docs/aws-guide.md](aws-guide.md)
**Monitoring:**
- **CloudWatch Logs:** Structured JSON logs from containers
- **CloudWatch Metrics:** Custom metrics (queue depths, processing times)
- **CloudWatch Alarms:** Alert on error rates, queue backlogs, high latency
- **ECS Service Metrics:** CPU, memory, task health
## Resilience Patterns
### 1. S3 Circuit Breakers
**Problem:** S3 throttling or outages can cascade to entire application
**Solution:** Circuit breakers fail fast, allow recovery
#### Implementation
Three independent circuit breakers:
- **`upload_circuit`** - For `put_object`, `upload_fileobj`
- **`download_circuit`** - For `get_object`, `head_object`, `list_objects`
- **`delete_circuit`** - For `delete_object`, `delete_objects`
#### Circuit States
1. **CLOSED** (Normal operation)
- Requests pass through to S3
- Track consecutive failures
- Threshold: 5 consecutive failures → OPEN
2. **OPEN** (Service degraded)
- Requests blocked immediately (fail fast)
- Raise `CircuitBreakerOpen` exception
- Timeout: 60 seconds → HALF_OPEN
3. **HALF_OPEN** (Testing recovery)
- Allow limited requests through
- 2 consecutive successes → CLOSED
- 1 failure → OPEN
#### Monitoring
Prometheus metrics:
```python
# Circuit breaker state (0=closed, 1=half-open, 2=open)
s3_circuit_breaker_state{circuit_name="s3-upload"} 0
# Operation results
s3_operations_total{operation="put_object", bucket="temp", result="success"} 1234
s3_operations_total{operation="put_object", bucket="temp", result="error"} 5
s3_operations_total{operation="put_object", bucket="temp", result="circuit_open"} 12
# Operation latency
s3_operation_duration_seconds{operation="put_object", bucket="temp"} 0.123
```
### 2. Exponential Backoff Retry
**Problem:** Transient errors (network blips, rate limits) fail operations unnecessarily
**Solution:** Retry with exponential backoff, categorize errors
#### Retry Logic
```python
max_attempts = 3
base_delay = 1.0 # seconds
max_delay = 8.0 # seconds
for attempt in range(1, max_attempts + 1):
try:
return s3_client.put_object(...)
except ClientError as e:
if is_retriable(e) and attempt < max_attempts:
delay = min(base_delay * (2 ** (attempt - 1)), max_delay)
sleep(delay)
continue
raise
```
#### Error Categorization
**Retriable errors** (temporary, retry):
- `503 Service Unavailable`
- `SlowDown` (rate limiting)
- `RequestTimeout`
- `InternalError`
- `ThrottlingException`
**Non-retriable errors** (permanent, fail immediately):
- `403 Forbidden` (permissions)
- `404 Not Found`
- `NoSuchKey`
- `AccessDenied`
#### Boto3 Adaptive Retry
Enabled in S3 client configuration:
```python
config = Config(
retries={
'mode': 'adaptive', # Intelligent client-side rate limiting
'max_attempts': 3
}
)
```
### 3. Rate Limiting (Redis Sliding Window)
**Problem:** Prevent abuse, ensure fair resource allocation
**Solution:** Redis-based sliding window rate limiter
#### Algorithm
```python
# Track requests in time-sorted list
key = f"rate_limit:{ip_address}"
now = time.time()
window = 60 # seconds
# Remove old requests outside window
redis.zremrangebyscore(key, 0, now - window)
# Count requests in current window
count = redis.zcard(key)
# Check limit
if count >= limit:
raise RateLimitExceeded("Too many requests")
# Add current request
redis.zadd(key, {request_id: now})
redis.expire(key, window)
```
#### Rate Limits
- **Per-IP:** 60 requests/minute
- **Global:** 1000 requests/minute
- **API endpoints:**
- `POST /submit`: 10/minute per IP
- `GET /status`: 120/minute per IP
- `GET /result`: 120/minute per IP
### 4. Graceful Shutdown
**Problem:** Killing workers mid-job causes incomplete processing
**Solution:** Shutdown event signals workers to finish current job
#### Implementation
```python
shutdown_event = asyncio.Event()
async def worker_loop(shutdown_event):
while not shutdown_event.is_set():
job = await queue.dequeue(timeout=5) # Short timeout for responsiveness
if job:
await process_job(job) # Complete current job
# On shutdown signal
shutdown_event.set()
await asyncio.wait_for(worker_task, timeout=30) # Max 30s grace period
```
#### Shutdown Sequence
1. Receive SIGTERM (Kubernetes/ECS termination)
2. Set shutdown event (workers stop accepting new jobs)
3. Wait for current jobs to complete (max 30 seconds)
4. Force cancel if timeout exceeded
5. Close Redis/S3 connections
6. Exit
## Architectural Decisions
### Why Monolith with Background Workers?
**Decision:** Single Python application with multiple worker threads
**Alternatives considered:** Microservices, separate worker processes
**Reasoning:**
- **Simplicity:** Shared codebase, single deployment, easier debugging
- **Shared state:** Direct access to Redis and S3 clients (no network overhead)
- **Development velocity:** Faster iteration, no inter-service communication
- **Resource efficiency:** Lower overhead than multiple containers
- **Sufficient scale:** UIC use case (<1000 documents/day, not Netflix-scale)
**Trade-offs:**
- Less flexible scaling (can't scale PII worker independently)
- Single point of failure (but ECS auto-restarts)
- Future migration path: Extract workers into separate services if needed
### Why Sorted Sets for Timeouts?
**Decision:** Redis sorted sets with Unix timestamps as scores
**Alternatives considered:** Background polling with `HGETALL` scans, separate timeout service
**Reasoning:**
- **Efficient queries:** `ZRANGEBYSCORE` finds expired in O(log N + M)
- **Automatic ordering:** No need to sort, Redis handles it
- **Atomic operations:** `ZADD`/`ZREM` prevent race conditions
- **Flexible deadlines:** Can extend timeout by updating score
**Trade-offs:**
- Requires scheduled worker (timeout worker runs every 30s)
- Not real-time (30s delay between timeout and action)
### Why Circuit Breakers for S3?
**Decision:** Independent circuit breakers for upload/download/delete
**Alternatives considered:** Simple retry, no circuit breaker, single shared circuit
**Reasoning:**
- **Fail fast:** Avoid waiting for timeouts when S3 is degraded
- **Cascading failure prevention:** Don't queue thousands of jobs when S3 is down
- **Independent failures:** Upload throttling shouldn't block downloads
- **Recovery testing:** Half-open state tests if S3 has recovered
**Trade-offs:**
- Complexity (state management, metrics)
- False positives (temporary failure could open circuit unnecessarily)
## Monitoring & Observability
### Health Checks
**Endpoint:** `GET /health`
```json
{
"status": "healthy",
"version": "0.1.0",
"checks": {
"redis": {"status": "ok", "latency_ms": 2},
"s3": {"status": "ok", "latency_ms": 45},
"workers": {
"pii_worker": "running",
"timeout_worker": "running"
}
},
"uptime_seconds": 3600
}
```
### Prometheus Metrics
**Queue metrics:**
```
queue_depth{queue="pii"} 5
queue_depth{queue="processing"} 2
```
**Job metrics:**
```
jobs_total{status="completed"} 1234
jobs_total{status="failed"} 12
jobs_total{status="denied"} 5
```
**Processing metrics:**
```
processing_duration_seconds{stage="pii_scan"} 15.3
processing_duration_seconds{stage="docling_conversion"} 45.2
processing_duration_seconds{stage="ai_enhancement"} 123.4
```
**S3 metrics:**
```
s3_operations_total{operation="upload", result="success"} 1234
s3_circuit_breaker_state{circuit="upload"} 0
s3_operation_duration_seconds{operation="upload"} 0.123
```
### Grafana Dashboards
**Overview dashboard:**
- Jobs submitted (rate per minute)
- Jobs completed (rate per minute)
- Success rate (completed / submitted)
- Queue depths (real-time)
- Worker health (up/down)
**Performance dashboard:**
- P50/P95/P99 processing times
- API latency distribution
- S3 operation latency
- Redis operation latency
**Errors dashboard:**
- Error rate by type
- Failed jobs by reason
- Circuit breaker trips
- Retry attempts
## Related Documentation
- **[Environment Setup](environment-setup-simplified.md)** - Local development setup guide
- **[AWS Deployment Guide](aws-guide.md)** - Production deployment on AWS ECS
- **[Rate Limiting](rate-limiting.md)** - Rate limiting implementation details
- **[CI/CD](ci-cd.md)** - GitHub Actions workflows and deployment automation
- **[CLAUDE.md](../CLAUDE.md)** - AI agent development guide and patterns
## Version History
- **v0.1.0** (2025-01-15) - Initial architecture document
- Complete file structure with all 50+ files
- Technology stack with versions
- Detailed processing pipeline (5 stages)
- Redis data structures deep dive
- Infrastructure comparison (local vs production)
- Resilience patterns (circuit breakers, retry, rate limiting)
- Merged content from program-flow.md
- Edge cases and error handling
- Performance characteristics
- Monitoring and observability