Large traces and performance

Perseval is designed to keep trace size out of the UI’s retained-object count.

• run pages: at most 200 rows;
• span pages: at most 500 rows;
• cached pages: at most 8 per list;
• UI delta application: at most 512 coalesced deltas per frame;
• finalized topology: persisted once rather than reconstructed during queries;
• child loading: lazy by direct parent;
• payload display: explicit and bounded.

Selection uses stable identities rather than vector indices. Loading a new batch does not reload the complete trace.

The current development suite has streamed and explored deterministic 20,000- and 100,000-span fixtures without an observed crash or swap. In the recorded 100,000-span run, peak physical footprint was approximately 200 MiB and the UI retained about 1,001 visible/cached rows rather than all spans.

Recorded live release-profile ingestion on the qualification machine was approximately:

Warm summary and page queries remained within the development acceptance budget.

The scale fixtures are deterministic expansions of real agent-trace shapes. They qualify paging, topology, storage, and responsiveness. They do not prove production prevalence, every vendor mapping, broken-trace reconstruction quality, or SpanLink behavior at the same scale.

• send retryable batches instead of one maximum-size request;
• keep attributes bounded and externalize payload-like content;
• preserve parent IDs and SpanLinks;
• use explicit project/build/session identity;
• watch queue, journal, projection, and analysis counters separately.

If a trace is slow to explore, narrow the active project/build/session scope, expand only relevant branches, and use search on loaded pages. Full Trace intentionally avoids an unbounded “load all spans” operation.