### caching2

```{eval-rst}
.. plantuml::

   @startuml
   title Cache Miss → Hydrate → AI Tables → Consumer Query (Minimal)
   
   
   skinparam sequence {
     ArrowColor black
     ArrowThickness 1
     ParticipantBorderColor black
     ParticipantBackgroundColor #FFFFFF
     LifeLineBorderColor black
     LifeLineBackgroundColor #FFFFFF
   }
   
   
   actor Consumer as C
   participant "Kafka Brokers" as K
   participant "Fast Cache (NVMe/SSD)" as FC
   participant "Cold Tier (Object Store)" as CT
   participant "AI Pipelines\n(Chunk/Embed, Feature Compute)" as AIP
   database "Lakehouse\n(AI Tables: Features, Embeddings, Labels)" as LH
   
   
   == Operational read ==
   C -> K : Subscribe / poll for records
   K -> FC : Read segment/key
   FC --> K : MISS
   FC -> CT : Hydrate segment/key (fetch)
   CT --> FC : Segment bytes
   FC --> K : Segment pointer / data (PROMOTE)
   K --> C : Deliver records (ms SLO)
   
   
   note over FC,CT
     Hydrate under SLO (e.g., p95 < 250 ms).
     Segment promoted to cache on success.
   end note
   
   
   == Stream → AI tables (parallel/continuous) ==
   K -> AIP : Stream events/docs
   AIP -> LH : Write/merge to AI Tables
   LH --> AIP : Ack / checkpoint
   
   
   note over AIP,LH
     Features, embeddings, labels
     materialized for training & RAG.
   end note
   
   
   == Analytics / AI batch reads ==
   C -> LH : BI/ML/RAG query on AI Tables
   LH --> C : Result set
   @enduml