Where the SDK's RPC calls really go

3 ใน 4 ของ call บน HyperEVM และ 1 ใน 3 บน BNB ยิงไปหา pool ที่ไม่มีอยู่จริงหรือไม่มีสภาพคล่อง — SDK สร้าง model ให้ทุกคู่ token × ทุก route โดยไม่เคยเช็คว่า pool ถูก deploy และไม่จำผลไว้ข้ามการ quote

Three quarters of the multicall inner calls on HyperEVM (one third on BNB) are spent on pools that were never created or have zero liquidity. The SDK instantiates a model for every token pair on every route without checking pool existence, and rebuilds everything from scratch on each quote. Fixing that takes HyperEVM from 26 requests per quote to 2 with byte-identical prices. Cutting correspondent tokens or routes blindly, the tempting fix, measurably worsens prices.

Every getQuote() does exactly one round of network I/O: it registers models, collects each model's state calls, packs them into Multicall2.tryAggregate batches, fires all batches at once, then computes routes purely in memory. So RPC cost is decided by two things only: how many models get registered and how they are batched. Split on/off, trade size, and gas price change nothing on the wire (verified).

pairs    = generateAllTokenPairs(src, dest)     HyperEVM 11 pairs (4 corr tokens) · BNB 28 pairs (8 corr tokens)
models   = Σ routes × pairs the route "supports"  every V2/V3 route claims EVERY pair → 289 / 914 models
calls    = Σ models × state calls                 V2: 1 · V3: 4 · Hybra V4: 5 · Solidly: 4 · Curve: 3 + n
requests = ceil(calls / maxCallsPerFetch) + 1     (+1 = a separate eth_call for dest.decimals())

Every pair on HyperEVM lands at 72–77 % waste; every BNB pair among the correspondent tokens resolves to the same 28-pair matrix and the same 36 %.

On HyperEVM, 350 of the 572 tick-querier calls (the heaviest call type, ~500 B each) go to pools that do not exist; tick data is 91 % of the 327 KB payload. Five routes never have a usable pool for any pair tested and cost 231 calls (19 %) per quote: hybra-v4-100, laminar-v3-10000, ultrasolid-v3-10000, prjx-v3-10000, hybra-v3-7500.

HyperEVM · requests per quote by maxCallsPerFetch

HyperEVM · fetch time (ms)

50 (today) → 200 cuts requests 3.6× at the same latency. A single 1,300-call batch reverts on the public node. Sequential batches at 50 take 7.7 s, which is why the SDK fires everything in parallel today.

BNB · fetch time (ms) by batch size

bsc-dataseed serves many small batches faster than a few big ones (100 → 1.06 s, 500 → 1.6–1.9 s, 1,000 → 4.5 s). The batch size must be a per-provider profile, not a per-chain constant.

Cutting intermediate tokens does slash RPC cost, but it removes exactly the hops that large trades need. Measured with the real getQuote():

• Constant data re-fetched every quote. BNB: Solidly/Thena decimals() × 112 and factory.getFee() × 56 — both immutable. Velodrome/Hybra V4 fee() × 58, Curve A/fee/totalSupply.
• dest.decimals() is a separate eth_call outside the multicall: +1 request per quote on a 100 req/min budget.
• Zero reuse across quotes. HyperEVM blocks every ~1 s, BNB every ~0.75 s; two quotes inside the same block re-download identical state.
• Parallelism is all-or-nothing. 25 batches via Promise.all with no concurrency cap, retry or back-off — this is what trips the public limiter (429).
• Code health. getReserves() exists only for the Uniswap V2/V3 families (35 model files omit it, so the liquidity validator would throw for Curve/Solidly/DODO/Balancer); dagre, @xyflow/react are unused runtime dependencies that pull React into every install; the bundle (80 KB gzip) ships all 7 chains' route tables and 55 ABIs to every consumer.

Projected end state (R1–R5, R8)

Do not: cut correspondent tokens (§5), disable split (no RPC saving), or replace Multicall with per-pool calls. Suggested order: R5 + R9(deps) + R10 as 1.1.1 → R1 + R2 + R3 as 1.2.0 → R4/R6/R8 → R7 + per-chain builds as 1.3.0.

After optimisation a quote needs 2–3 eth_call batches of ≤ 200–500 calls. Choose on batch tolerance, requests/second and HyperEVM support; archive/trace/WebSocket do not matter here.

• benchmark/rpc-profile.ts rebuilds exactly what MoroBestRate._fetch() registers, executes the same batches itself, and attributes every inner call to route / model / selector, classifying each model as ok, dead (0 liquidity), empty (no contract), or failed (revert). Dead is only detectable for the Uniswap V2/V3 families (see §6).
• benchmark/quote-quality.ts runs the real getQuote() at 1 / 100 / 1,000 units under different configurations and compares amountOut, requests and time.
• Public RPCs, same day, same block range per comparison. The RPC panel under every quote on the Moro Swap page uses the same counting technique live.

benchmark/profile.sh --chain hyperevm --src HYPE --dest USDT0 --live-only
benchmark/profile.sh --chain bsc --batch 100 --corr 4 --json out.json
npx ts-node -T -O '{"module":"commonjs"}' benchmark/quote-quality.ts --chain hyperevm --drop hybra-v3-7500

Full write-up with every table: INVESTIGATE.md on branch feat/optimize at korrio/moro-sdk-optimized. Production SDK on npm is unchanged: @moromoro/moro-sdk.

Each optimisation round is measured by benchmark/ab-test.ts: the published npm SDK (A) and the candidate build (B) run the identical workload (3 pairs × 1 / 100 / 1,000 units per chain) with every eth_call pinned to the same block, interleaved with alternating order, N repetitions, medians reported. The gate is on price: B fails if its amountOut is worse than A's by more than 5 bps on any case. Cost improvements are reported, never assumed.

Protocol, metric definitions, per-round template, and pros/cons of R1–R10: benchmark/AB-TESTING.md in the optimize repo. Results of each round are committed under benchmark/results/ and summarised as Report #2, #3, … on this site.