zhanghongqing/knowledge-log is a Flink ETL pipeline that builds relationship graphs from DNS and connection logs (Ip2IpGraphProcessFunction, DnsGraphProcessFunction/DnsRelationProcessFunction) and persists them to ArangoDB (a graph database) plus ClickHouse, indicating infrastructure for graph-based flow-correlation analysis (e.g. clustering related proxy/VPN infrastructure by IP-to-IP or DNS relationships) separate from the per-packet DPI engine.

relationship-gtpc-user derives a single per-subscriber correlation key as Hashkey = md5(IMEI + IMSI + phone_number + vsys_id) from mobile-core GTP-C signaling, and a separate per-tunnel key gtp_uuid = md5(uplink_teid + "|" + downlink_teid). On every update it writes not just the primary user<->TEID row (tsg_galaxy:relation_user_teid) but up to four additional secondary-index rows into tsg_galaxy:gtpc_knowledge_base, one per identifier (IMEI, IMSI, phone number, APN) with a type-digit prefix ("0"/"1"/"2"/"3") and — for phone number and APN — the value's characters reversed before use as the row key, a range-scan-friendly key design that lets any of the four identifiers be used to look up the associated tunnel/session independently.

radius-relationship-hbase (galaxy/tsg_olap) builds a bidirectional, MD5-keyed RADIUS accounting index in HBase — a framed-IP-to-account table and its account-to-IP mirror — so any subscriber account or assigned IP can be looked up in O(1) from either direction, per multi-tenant vsys_id. Each direction is fronted by a Guava write-through cache (10M-entry cap, 24h TTL) so lookups do not hit HBase on the hot path.

radius-account-knowledge's MapCompletedFunction explicitly prefers the RADIUS acct_multi_session_id field over acct_session_id when both are present in an accounting record, specifically to keep a single logical session identifier stable across events that would otherwise fragment it (e.g. re-authentication, mobility/roaming, NAS handoff) — a deliberate design choice for session-identity continuity rather than a default library behavior.

The VoIP-fusion correlation key linking a SIP signaling record to its RTP media record is not timing/heuristic-based -- it is an exact match between the negotiated media endpoint fields parsed out of the SIP record (from_rtp_ip/from_rtp_port/to_rtp_ip/to_rtp_port, i.e. the SDP-negotiated RTP endpoints carried in the SIP signaling) and the actual observed RTP flow's four-tuple (common_client_ip/port, common_server_ip/port). A separate normalization rule sorts every four-tuple by port then IP (smaller first) before use as a join key, and common_vsys_id is the default grouping/partition key.

VoipCalibrationWindowFunction only auto-merges a candidate group when it is exactly one-to-many (1 RTP matched to N SIP records, or 1 SIP matched to N RTP records); any other cardinality within a correlation-key bucket (e.g. 2 SIP + 2 RTP) is explicitly logged as anomalous ("Not logical") and every record in that bucket is emitted individually as an unmerged "violation" log rather than being forced into a merge or silently dropped.

VoIP correlation runs in two sequential time windows, not one -- an initial tumbling window (voip.calibration.window.time, configured at 30s) attempts to fuse SIP/RTP records sharing a key, and any record still unmatched after that window is held in a second in-memory cache and retried against a longer secondary window (sec.combine.sr.cache.secs, configured at 60s) before finally being emitted unmerged -- giving the pipeline up to ~90s of tolerance for SIP and RTP legs of the same call arriving out of sync.