Optimisation round 2: one shared memory for every SDK
1.3.0 ให้ทุก process ของ SDK แชร์สิ่งที่เรียนรู้ผ่าน Redis — pool ไหนไม่มี, pool ไหนว่างเปล่า, token มีกี่ทศนิยม — process ใหม่จึงเริ่มด้วยความรู้ของทุกคนก่อนหน้า ผล: call ลดอีก 25–30 % บน HyperEVM ตั้งแต่ quote แรก, request บน BNB ลด 20 % ราคาเท่าเดิม 0.00 bps ทั้ง 18 กรณี
Round 1 stopped asking about pools that do not exist. Round 2 adds two things: the SDK now notices pools that exist but are empty (a quarter of what was left on HyperEVM) and skips them for an hour, and every SDK process publishes what it learned to a shared Redis so the next process, on any machine, starts warm. Measured with the same A/B protocol against the frozen 1.2.0 build. Δ price 0.00 bps on all 18 cases.
1 · Three versions, one trade (1 HYPE → USDT0, first quote of a fresh process)
2 · A/B: frozen 1.2.0 build vs 1.3.0, shared cache warmed by another process
| chain · pair | requests | inner calls | bytes | Δ price |
|---|---|---|---|---|
| HyperEVM · HYPE → USDT0 | 2 → 2 | 432 → 321 (−26 %) | −26 % | 0.00 bps |
| HyperEVM · USDT0 → WHYPE | 2 → 2 | 401 → 297 (−26 %) | −26 % | 0.00 bps |
| HyperEVM · kHYPE → USDT0 | 3 → 2 (−33 %) | 568 → 396 (−30 %) | −31 % | 0.00 bps |
| BNB · BNB → USDT | 5 → 4 (−20 %) | 1,351 → 1,190 (−12 %) | −17 % | 0.00 bps |
| BNB · BUSD → USDT | 5 → 4 (−20 %) | 1,351 → 1,190 (−12 %) | −17 % | 0.00 bps |
| BNB · ETH → CAKE | 5 → 4 (−20 %) | 1,351 → 1,190 (−12 %) | −17 % | 0.00 bps |
Same route chosen in every case. The first case of each run carries the one-time Redis connect + sync (~1.2 s), which shows up as +14–18 % wall time on that case only; every later quote is 4–31 % faster than 1.2.0.
3 · Why the cache holds facts, not pool state
Before designing the feature we measured the Redis instance from the SDK's point of view:
| operation | measured | consequence |
|---|---|---|
| connect | 651 ms | paid once per process; started in the constructor so the first quote rarely waits the full amount |
PING round trip | 332 ms | a quote must never wait on Redis → sync once at start (bounded 2 s), then every 30 s in the background |
300 KB SET + GET | 3,037 ms | slower than fetching the same pool state from the RPC (≈1.0 s) → pool state is not shared; only skip-lists and decimals (a few KB) |
4 · Trade-offs and what to fix before a public release
- The Redis credential was built into the SDK in 1.3.0–1.3.4 (by request), so anyone with the package could read and write the cache. Removed in 1.3.5: the URL now comes from
MORO_SHARED_CACHE_REDIS_URLon the server (orsharedCache.redisUrl), the cache is off otherwise, and the leaked credential must be rotated. A malicious write could hide live pools (worse prices; never a wrong swap, because the router still enforcesminAmountOut). Before publishing: rotate it, give the SDK a read-only ACL user, and write only from trusted servers or a small proxy. - Node only. Browsers cannot open TCP connections; Moro Swap keeps its in-memory caches and shows "shared cache: not reachable from a browser". A read-only HTTP proxy in front of Redis (
sharedCache.httpUrl) brings the same −25 % to the web app; that is the next step. - First quote per process waits up to 2 s for the initial sync. Right for servers and bots; for one-shot CLIs pass
{ sharedCache: false }. - The instance is far from the users (330 ms). Closer Redis (or an edge KV) would make a block-scoped shared state cache viable later.
- What is left: tick data is now 88 % of the bytes → adaptive tick window (R6); per-chain builds (R9); a scheduled registry refresh that writes straight into the shared cache.
5 · Reproduce
# warm the shared cache from one process, then A/B a fresh process against the frozen 1.2.0 build
npx ts-node -T -O '{"module":"commonjs"}' benchmark/quote-quality.ts --chain hyperevm
npx ts-node -T -O '{"module":"commonjs"}' benchmark/ab-test.ts --chain hyperevm --a v1.2.0 --b dist --runs 2 --pause 15000
npx ts-node -T -O '{"module":"commonjs"}' benchmark/ab-test.ts --chain bsc --a v1.2.0 --b dist --runs 2
Write-up: benchmark/rounds/2.md; raw samples: benchmark/results/2026-09-13-07-0*-round2-v1.3.0.json; code: src/services/sharedCache.ts — on korrio/moro-sdk-optimized. The npm package stays at 1.1.0 until the team publishes.