schema_version: 1 case_id: ticdc-dataflow-stall-during-incident-window ground_truth: root_cause: TiCDC was stalled after puller because the Kafka sink path was repeatedly failing with EOF, broken-pipe, and TLS close-notify errors against the downstream Kafka proxy, so checkpoint lag kept growing while upstream activity remained present. judge: semantic_strategy: undecided contract: checks: - id: time_window_match required: true description: Actual diagnosis time window matches one allowed case window. match: all_of rules: - source: trace_fact matcher: bool_true path: time_window_match - id: cluster_scope_match required: true description: Actual diagnosis cluster scope matches expected cluster id. match: all_of rules: - source: trace_fact matcher: bool_true path: cluster_scope_match workflow: checks: - id: architecture_context_loaded weight: 5 required: false description: Architecture context is loaded. match: all_of rules: - source: workflow_fact matcher: bool_true path: architecture_context_loaded - id: misdiagnosis_guardrails_loaded weight: 5 required: true description: Misdiagnosis guardrails are loaded. match: all_of rules: - source: workflow_fact matcher: bool_true path: misdiagnosis_guardrails_loaded - id: diagnosis_run_initialized weight: 5 required: true description: Diagnosis run is initialized. match: all_of rules: - source: workflow_fact matcher: bool_true path: diagnosis_run_initialized - id: evidence_collection_completed weight: 10 required: true description: Evidence collection completes. match: all_of rules: - source: workflow_fact matcher: bool_true path: evidence_collection_completed - id: summary_read_completed weight: 5 required: false description: Summary artifact is read when available. match: all_of rules: - source: workflow_fact matcher: bool_true path: summary_read_completed - id: final_answer_emitted weight: 0 required: false description: Final answer is emitted. match: all_of rules: - source: workflow_fact matcher: bool_true path: final_answer_emitted execution_contract: checks: - id: required_workflow_present required: true description: Required workflow milestones are present. match: all_of rules: - source: workflow_check matcher: subset_observed paths: - misdiagnosis_guardrails_loaded - diagnosis_run_initialized - evidence_collection_completed - id: required_artifacts_present required: true description: Required diagnosis report artifact is present. match: all_of rules: - source: artifact_fact matcher: subset_observed paths: - diagnosis_report_present - id: preferred_channels_minimum_observed required: true description: At least one preferred evidence channel is observed. match: all_of rules: - source: channel_fact matcher: count_gte paths: - logs - query_or_metric threshold: 1 - id: preferred_channels_all_observed required: false description: Preferred evidence channels are all observed. match: all_of rules: - source: channel_fact matcher: subset_observed paths: - logs - query_or_metric evidence_collection: weight: 5 checks: - id: required_artifacts_present weight: 4 required: true description: Required diagnosis report artifact is present and readable. match: all_of rules: - source: artifact_fact matcher: subset_observed paths: - diagnosis_report_present - id: preferred_channels_minimum_observed weight: 0 required: true description: At least the minimum number of preferred evidence channels are observed. match: all_of rules: - source: channel_fact matcher: count_gte paths: - logs - query_or_metric threshold: 1 - id: preferred_channels_all_observed weight: 6 required: false description: Preferred evidence channels are all observed. match: all_of rules: - source: channel_fact matcher: subset_observed paths: - logs - query_or_metric critical_evidence: weight: 30 checks: - id: primary_symptom_evidence_present weight: 8 description: Direct TiCDC downstream-stall evidence is present within the incident window. match: all_of rules: - source: logs matcher: keyword_any keywords: - "schedulerv3: slow table" - source: logs matcher: keyword_any keywords: - "checkpointLag=" - source: logs matcher: keyword_any keywords: - "changefeed=changefeed-2046429488116162560" - id: root_cause_mechanism_evidence_present weight: 10 description: Direct Kafka sink connectivity failure evidence is present within the incident window. match: all_of rules: - source: logs matcher: keyword_any keywords: - "client/metadata got error from broker" - "Closed connection to broker" - source: logs matcher: keyword_any keywords: - "kafka-sink-proxy" - "write: broken pipe" - "failed to send closeNotify alert" - id: changefeed_context_evidence_present weight: 6 required: false description: Evidence includes TiCDC changefeed-specific log context rather than only generic infrastructure logs. match: any_of rules: - source: logs matcher: keyword_any keywords: - "changefeed=changefeed-2046429488116162560" - "ticdc-1379661944646414117-ticdc-" evidence_routing: weight: 5 checks: - id: analysis_patch_valid_json weight: 1 required: true description: Analysis patch is valid JSON whenever the patch exists. match: any_of rules: - source: routing_fact matcher: bool_false path: analysis_patch_present - source: routing_fact matcher: bool_true path: analysis_patch_valid_json - id: analysis_items_shape_valid weight: 2 required: true description: Analysis patch items are structurally valid whenever the patch exists. match: any_of rules: - source: routing_fact matcher: bool_false path: analysis_patch_present - source: routing_fact matcher: bool_true path: analysis_items_shape_valid - id: analysis_key_evidence_present weight: 2 required: true description: Analysis patch includes key evidence whenever the patch exists. match: any_of rules: - source: routing_fact matcher: bool_false path: analysis_patch_present - source: routing_fact matcher: bool_true path: analysis_has_key_evidence - id: metric_refs_resolvable weight: 1 required: true description: Metric references are resolvable whenever the patch contains metric refs. match: any_of rules: - source: routing_fact matcher: bool_false path: analysis_metric_refs_required - source: routing_fact matcher: bool_true path: analysis_metric_refs_resolvable - id: query_refs_resolvable weight: 1 required: true description: Query references are resolvable whenever the patch contains query refs. match: any_of rules: - source: routing_fact matcher: bool_false path: analysis_query_refs_required - source: routing_fact matcher: bool_true path: analysis_query_refs_resolvable - id: supporting_evidence_present weight: 0 required: false description: Analysis patch includes supporting evidence whenever the patch exists. match: any_of rules: - source: routing_fact matcher: bool_false path: analysis_patch_present - source: routing_fact matcher: bool_true path: analysis_has_supporting_evidence - id: analysis_log_or_event_channel_resolvable weight: 0 required: false description: Routed log or event items stay resolvable against observed channels. match: all_of rules: - source: routing_fact matcher: bool_true path: analysis_log_or_event_channel_resolvable diagnosis: checks: - id: identify_downstream_pipeline_stall_after_puller mode: blended weight: 3 required: true description: Identifies the primary problem as a downstream TiCDC pipeline stall after puller rather than a source-side no-data condition. deterministic: match: any_of rules: - source: output_text matcher: keyword_any keywords: - after puller - post-puller - downstream blocked - downstream pipeline - downstream stages - sorter -> mounter -> sink - sink 完全阻塞 - puller 正常工作 - "puller: ok" - "upstream tikv → [puller" - upstream data flows - lock resolver events confirm - puller still has data - puller still advancing - puller is not the bottleneck - 下游阶段 - 下游处理链路 - 后续阶段 - puller 还有数据 - puller 仍在推进 - 不是 puller 问题 - 不是 puller 本身 - 不是puller本身 semantic: id: 1 title: 主路径判断 description: | 判断回答是否识别主问题发生在 puller 之后的下游处理阶段， 而不是把现象误判成 source side 没数据或 puller 本身故障。 - id: identify_kafka_sink_connectivity_primary_mechanism mode: blended weight: 3 required: true description: Identifies Kafka sink connectivity or proxy-side connection failure as the primary blocking mechanism. deterministic: match: any_of rules: - source: output_text matcher: keyword_any keywords: - Kafka Sink - kafka sink - kafka proxy - broker connectivity - metadata got error from broker - broken pipe - closeNotify - EOF - sarama - sink 完全阻塞 - Kafka 代理 - Kafka 连接 - broker 连通性 semantic: id: 2 title: 机制定位 description: | 判断回答是否把主机制收敛到 Kafka sink 侧的连接失败、broker/proxy 异常、 EOF / broken pipe / TLS 相关错误，而不是停留在泛泛的“CDC 慢”。 - id: reject_puller_or_source_side_primary_root_cause mode: deterministic_only scoring_mode: penalty weight: 3 required: false description: Does not misdiagnose the incident as a puller-side or source-side primary root cause. deterministic: match: all_of rules: - source: output_fact matcher: bool_false path: puller_or_source_side_primary_root_cause_asserted - id: use_checkpoint_or_dispatcher_lag_as_supporting_evidence mode: blended weight: 3 required: false description: Uses checkpoint lag or frozen checkpoint progression as supporting evidence for the downstream stall. deterministic: match: any_of rules: - source: output_text matcher: keyword_any keywords: - owner_ckp_lag - checkpoint_ts - checkpoint lag - checkpointTs - checkpoint not advancing - lag growing at - 1.000s/s - 线性增长 - checkpoint 延迟 - checkpoint 没有推进 - checkpoint 没推进 semantic: id: 3 title: 佐证使用 description: | 判断回答是否把 checkpoint lag 持续增长、checkpoint_ts 冻结或类似的 时序信号当作支持性证据，用来证明问题确实卡在 puller 之后的处理链路。 - id: recommend_kafka_sink_connectivity_followup mode: blended weight: 3 required: false description: Recommends follow-up actions aligned with Kafka sink connectivity, proxy, broker, or TLS investigation. deterministic: match: any_of rules: - source: output_text matcher: keyword_any keywords: - inspect kafka proxy - check broker connectivity - check TLS - check certificate - check kafka broker logs - 排查 Kafka 代理 - 排查 broker 连通性 - 检查 TLS - 检查证书 - 检查 Kafka broker 日志 semantic: id: 4 title: 修复建议对齐 description: | 判断回答是否把后续动作对齐到 Kafka sink / proxy / broker / TLS 侧排查， 而不是只给笼统的扩容建议。