Runtime Cost Model

May 26, 2026 ยท View on GitHub

This page gives an engineering estimate of the storage and write amplification created by Aetheris durable execution. Exact byte size depends on payload size, metadata, tracing detail, and database encoding.

Baseline Event Counts

Step typeTypical recordsNotes
Job submission2-3 eventsjob_created, plan_generated, optional job_queued
Pure/LLM node3-5 eventsnode_started, command_emitted, command_committed, node_finished, optional step_committed
Runtime Tool node5-8 eventsnode events plus tool_invocation_started, tool_invocation_finished, command commit, summaries
Wait / HITL node2-4 events before resumenode_started, job_waiting / job_parked; later wait_completed / job_resumed
Job completion1-2 eventsjob_completed, optional checkpoint/session updates

Write Amplification by Store

StoreWhen writtenPurpose
Job/Event StoreEvery lifecycle transition and execution eventReplay authority, lease state, audit base
Checkpoint StoreStep or checkpoint boundaryResume cursor/state without replaying long histories
ToolInvocationStoreRuntime Tool acquire/commitInvocation Ledger arbitration
Effect StoreCompleted Tool/LLM/HTTP effectsStrong Replay catch-up and effect metadata
Trace/Audit derived viewsOptional or API-derivedUI, export, forensics

Example: 10-Step Agent Job

Assume:

  • 3 LLM nodes
  • 5 Runtime Tool nodes
  • 1 wait/resume node
  • 1 pure/workflow node

Approximate writes:

CategoryEstimate
Initial job/plan events2-3
LLM node events9-15
Tool node events25-40
Wait/resume events3-5
Pure/workflow node events3-5
Completion events1-2
Checkpointsup to 10
Effect Store recordsabout 8 (LLM + Tool effects)
ToolInvocationStore recordsabout 5

Expected total: roughly 45-70 event-store writes plus checkpoint/effect/ledger writes. Large LLM outputs or tool payloads dominate storage size more than event count.

Cost Controls

ControlEffect
SnapshottingReduces replay time for long event histories
Runtime GCRemoves expired checkpoints/tool invocation records according to retention policy
Event taxonomyKeeps replay-critical history separate from trace/audit expansion
Payload disciplineStore summaries in trace events and full payload only where replay/audit requires it
External HTTP migrationStart with one outer call, then extract only high-risk side effects into Runtime Tools

Sizing Questions

Before production rollout, answer:

  • Average and P95 steps per Job?
  • Ratio of LLM nodes to Tool nodes?
  • Average LLM/tool payload size?
  • Retention requirement for event history, effects, checkpoints, and audit exports?
  • Expected concurrent waiting Jobs?
  • Required replay latency for incident response?