LAB_STATUS: ONLINE/ACTIVE_EXPERIMENTS: 07/LAST_COMMIT: 2026-05-29 09:14 UTC/UPTIME: 99.98%/REGION: EDGE_GLOBAL
MERCTECHSResearch
Research Log / FIELD_NOTES

Field Notes

Benchmarks, post-mortems, and deep technical essays from the Merctechs engineering collective. We publish methodology and raw numbers — not conclusions you have to take on faith.

◆ Featured · NOTE_0042

Sub-100ms Cold Starts: Profiling Our Edge Isolate Pool

How we cut cold-start latency from 340ms to 94ms by rethinking isolate warm-up, V8 snapshot reuse, and the allocator behavior nobody talks about.

runtime
Read the note
FIG_01 // cold_start_latency_msn=2,400
2026-05-28
NOTE_0042
14 min read
runtime

Sub-100ms Cold Starts: Profiling Our Edge Isolate Pool

How we cut cold-start latency from 340ms to 94ms by rethinking isolate warm-up, V8 snapshot reuse, and the allocator behavior nobody talks about.

Read
2026-05-19
NOTE_0041
18 min read
ml

Distilling a 7B Model to Run On-Device Without Melting the Phone

A field report on 4-bit quantization, knowledge distillation, and the thermal budget that quietly governs every on-device inference decision.

Read
2026-05-08
NOTE_0040
22 min read
distributed

The Cost of Consistency: CRDTs vs Consensus in Production

We ran the same collaborative-editing workload through both a Raft cluster and a CRDT layer. The trade-offs were not where we expected them.

Read
2026-04-29
NOTE_0039
16 min read
data

Rethinking Backpressure in Event-Driven Pipelines

When a downstream consumer stalls, naive buffering turns a hiccup into an outage. Here's the credit-based scheme we landed on instead.

Read
2026-04-21
NOTE_0038
12 min read
ml

Vector Search at the Edge: Quantization Trade-offs

Product quantization buys you 8× smaller indexes — at a recall cost. We mapped exactly where that curve bends for sub-millisecond retrieval.

Read
2026-04-12
NOTE_0037
20 min read
security

WASM as a Plugin Sandbox: A Field Report

We let untrusted third-party code run in production behind a WebAssembly boundary. Here's the host-call overhead, the memory model, and the footguns.

Read
2026-03-30
NOTE_0036
15 min read
data

Zero-Downtime Schema Migrations at Scale

A dual-write and backfill harness that lets us evolve Postgres schemas under live traffic — with a rollback path we actually trust.

Read
2026-03-18
NOTE_0035
11 min read
runtime

The Allocator Is Lying to Your Latency Graphs

Tail latency that looked like network jitter turned out to be allocator fragmentation. How we found it, and the arena strategy that fixed it.

Read
2026-03-05
NOTE_0034
13 min read
tooling

Distributed Tracing Without the Sampling Regret

Head-based sampling drops the traces you actually need. We moved to tail-based sampling and measured what it costs to keep the interesting 0.1%.

Read
2026-02-20
NOTE_0033
17 min read
ml

Speculative Decoding: Measuring Real-World Latency Gains

The papers promise 2–3× throughput. We benchmarked draft-model speculation across batch sizes on commodity GPUs to find where the promise holds.

Read
Lab / RESEARCH_HOME

See what's running on the bench.

Live experiments, focus domains, and the collective behind the notes.