Epic: Meet the Phase 3 targets #1147

RPC v2 has to meet the newest Phase 3 targets, which the leads set on 2026-10-21. They are recorded in stellar-experimental/stellar-rpc-benchmarks docs/targets.json.

A ledger closes every 900 ms. It carries up to 9,000 SAC transfers, 5,400 OZ token transfers or 2,700 Soroswap swaps. A client has to get its transaction's result within 2,000 ms at p99, split like this:

sequenceDiagram
    participant Client
    participant RPC
    participant Core
    Client->>+RPC: send tx · 25 ms (network)
    RPC->>+Core: submit tx
    Core->>RPC: response · 10 ms p99 (RPC)
    deactivate RPC
    RPC-->>Client: submit ack, not counted because core is already working
    Note over Core: externalize + emit meta<br/>1,800 ms p99 (Core)
    Core->>-RPC: meta stream
    activate RPC
    Note over RPC: ingest · 105 ms p99 (RPC)
    deactivate RPC
    Client->>+RPC: getTransaction · 50 ms round trip (network)
    RPC->>-Client: result · 10 ms p99 (RPC)

In total, 25 + 10 + 1,800 + 105 + 50 + 10 = 2,000 ms. The network times assume a 50 ms round trip. The response to sendTransaction isn't counted, because core is already working on the transaction by then. Core owns the 1,800 ms. RPC owns three parts:

• Ingesting each ledger: 105 ms at p99.
• Handling sendTransaction: 10 ms at p99.
• Handling getTransaction: 10 ms at p99, while serving 1,000 requests per second for SAC ledgers, 600 for OZ token and 300 for Soroswap.

Each read endpoint also has its own SLA, in docs/sla-derivation.md. For getTransaction it is 20 ms at p99 on hot data, which isn't frozen yet, and 30 ms on cold data, at 300 requests per second.

Where we are

The last benchmark campaign ran on 2026-10-23. It tested commit c772750, from the pending PRs for the GA release (#1087 up to #1126). Among other things, those PRs stop hot ingestion from holding events in memory, which fixed the out-of-memory failures that stopped earlier runs. It ran once, on a c6id.8xlarge with local NVMe, over 10,000 sac-9000 ledgers at one every 900 ms, with diagnostic events still in the ledgers.

• Hot ingestion took 227.6 ms per ledger at p99 (163.0 ms at p50), against the 105 ms target.
• At 1,000 requests per second, getTransaction took 847 ms at p99 on hot data and 932 ms on cold data, against the 10 ms target.
• At 300 requests per second, getTransaction took 42.9 ms at p99 on hot data and 46.5 ms on cold data, against the SLA's 20 ms and 30 ms.

Sub-issues

In priority order:

Related

#905 and #906 are the proofs of concept behind #1151 and #1153.

🤖 Generated with Claude Code