Concurrent submissions

This commit is contained in:
stopflock
2025-12-23 17:56:16 -06:00
parent 8b44b3abf5
commit 607ecbafaf
5 changed files with 119 additions and 58 deletions

View File

@@ -202,15 +202,24 @@ Deletions don't need position dragging or tag editing - they just need confirmat
- Retries: Exponential backoff up to 59 minutes
- Failures: OSM auto-closes after 60 minutes, so we eventually give up
**Queue processing workflow:**
**Queue processing workflow (v2.3.0+ concurrent processing):**
1. User action (add/edit/delete) → `PendingUpload` created with `UploadState.pending`
2. Immediate visual feedback (cache updated with temp markers)
3. Background uploader processes queue when online:
3. Background uploader starts new uploads every 5 seconds (configurable via `kUploadQueueProcessingInterval`):
- **Concurrency limit**: Maximum 5 uploads processing simultaneously (`kMaxConcurrentUploads`)
- **Individual lifecycles**: Each upload processes through all three stages independently
- **Timer role**: Only used to start new pending uploads, not control stage progression
4. Each upload processes through stages without waiting for other uploads:
- **Pending** → Create changeset → **CreatingChangeset****Uploading**
- **Uploading** → Upload node → **ClosingChangeset**
- **ClosingChangeset** → Close changeset → **Complete**
4. Success → cache updated with real data, temp markers removed
5. Failures → appropriate retry logic based on which stage failed
5. Success → cache updated with real data, temp markers removed
6. Failures → appropriate retry logic based on which stage failed
**Performance improvement (v2.3.0):**
- **Before**: Sequential processing with 10-second delays between each stage of each upload
- **After**: Concurrent processing with uploads completing in 10-30 seconds regardless of queue size
- **User benefit**: 3-5x faster upload processing for users with good internet connections
**Why three explicit stages:**
The previous implementation conflated changeset creation + node operation as one step, making error handling unclear. The new approach: