Files
kefu/deploy/knowledge-batch-20261007/server-bundle/release-plan.json
T
2026-11-16 09:39:42 +08:00

57 lines
2.7 KiB
JSON

{
"release": "knowledge-batch-20261007",
"candidateRoot": "C:/kefu",
"backendFiles": [
"wechat_rpa/knowledge_api.py",
"wechat_rpa/knowledge_store.py",
"wechat_rpa/knowledge_worker.py",
"wechat_rpa/knowledge_models.py",
"wechat_rpa/knowledge_model_output.py"
],
"mergeBase": "production",
"frontendDir": "admin-web/apps/web-antd/dist",
"frontendSource": "C:/wechat_rpa/admin-web/apps/web-antd/dist",
"frontendHeadBuild": "frontend-head-build",
"frontendReplaced": [
"index.html"
],
"frontendAdded": [
"js/about-BbRvvX_R.js",
"js/bootstrap-DY4Zvm_d.js",
"js/center-DqZcbhRF.js",
"js/permissions-DaAFwVE9.js",
"js/routes-D3v9qRvJ.js",
"js/run-dialog-Bf3phWnn.js",
"js/run-format-DCPS1JEP.js",
"js/runs-panel-BEh9vfHE.js",
"js/settings-dialog-DDzO686S.js",
"js/wash-options-BlNWMMwm.js",
"jse/index-index-WE0ld32x.js"
],
"frontendNewPages": [
"js/run-dialog.js",
"js/runs-panel.js",
"js/wash-options.js"
],
"frontendRemovedPages": [],
"serverTestModules": [
"test_knowledge_runs",
"test_knowledge",
"test_knowledge_tasks",
"test_knowledge_model_output",
"test_knowledge_review",
"test_knowledge_search"
],
"imageTestModules": [
"test_archive_api"
],
"schemaMigration": "none (knowledge_config gains an optional model_batch value only when saved)",
"restartServices": [
"api",
"gateway",
"knowledge-worker"
],
"requeueTimedOutJobs": true,
"notes": "Hotfix on top of knowledge-wash-20261007 after the first production run (809,209 messages, doubao-seed-2.1-pro, 4-way): one account job stopped on '模型请求超时' and throughput was ~1 Q&A per model call with a thinking model. Changes: thinking disabled for Ark/DeepSeek/doubao models (direct request, falls back to the gateway path when the provider rejects the parameter); several Q&As per request (runs default 8, legacy jobs stay at 1; truncated/missing items retried one by one); timeouts, 429, 5xx and network errors back off 30s..10min and retry automatically up to 6 times instead of failing the job (a timed-out batch is retried one by one); model jobs read up to 25 pages per step to fill concurrency x batch; resume clears the backoff; UI: per-request batch size, speed/ETA, 'continue failed accounts' while others run. Backend files = kefu working tree, which descends from production (= knowledge-wash payload); merge base is production. Frontend head build = current tree with this release's frontend files reverted (= live), page graph equal except build time. 03-deploy re-queues batch jobs that had failed with the timeout message, since the new worker would have retried them."
}