HYX

Study 07 / Edge observability

Hyx research study

One request. One attributable record.

A schema and ingestion study of delivery, transport, security, signals, model, and fleet context written together at the edge.

Research classification

66-field schema / 16 focused tests / host-local serialization benchmark

E1 / finding
66
joined dimensions
one LogRow schema / current repository
E2 / finding
584.3ns
JSONEachRow encode
host release benchmark
E3 / finding
1.71M
rows / second
serialization only / single benchmark loop
E4 / finding
16
focused tests passed
2 benchmarks ignored by default

Abstract

The claim begins with a question, not a headline.

Research question / RQ-07

Can the edge preserve decision attribution in one queryable row while keeping serialization and ingestion work bounded?

Investigation slows down when identity, enforcement, performance, and model evidence live in different systems. Hyx writes the request context as one wide record, then batches it without allowing sampling to erase mandatory security outcomes.

01

Method

Inventory the complete row, trace sampling and must-log decisions into the batching worker, verify schema/encoding tests, and run the release serialization benchmark.

02

Finding

Delivery, security, Hyx AI, server, and network details share one timestamped request record.

03

Disclosure boundary

The public study describes schema families and ingestion guarantees without publishing customer identifiers, private reason taxonomies, or operational endpoints.

System boundary

Context is joined while the request still owns it.

S-01 / system topology

request / classification / action / fleet

01

Delivery

Status, bytes, cache state, origin latency, protocol, POP, and configuration version explain how the request was served.

02

Identity

IP-derived geography, ASN, TLS, JA4, header order, and bot evidence describe the client and the uncertainty around it.

03

Enforcement

WAF rule, action, anomaly score, named signals, model version, probability, latency, and reasons explain why policy resolved as it did.

04

Ingestion

Bounded channels, batch thresholds, timed flushes, schema verification, sampling, and must-log overrides govern movement into ClickHouse.

Controlled measurement

The local encoder clears a million rows per second; storage is outside the result.

M-01 / host-local evidence

release mode / scoped result

MeasurementObservedScope
01JSONEachRow encoding584.3ns / row1.71M rows / second
02Hand-rolled timestamp36.8nsbenchmark alternative
03Single timestamp format124.5nslibrary formatting path
04Double timestamp format150.1nsbaseline comparison
05Focused suite16 passedschema + encoding + sampling behavior

Release benchmark on this host. Measures JSONEachRow construction, not channel contention, network transfer, ClickHouse insertion, storage, or query latency.

Findings

Attribution is cheaper when it is not reconstructed later.

01

One causal surface

A connection can be followed through delivery and enforcement without a request-time join across independent telemetry products.

02

Schema parity is tested

The row definition, ClickHouse schema, and JSON encoding are checked together so additions cannot silently shift columns.

03

Sampling has exceptions

Routine traffic may be sampled, while security and operationally significant outcomes can force a record.

04

Batches isolate the request

Request processing sends a completed record into a bounded worker rather than performing a synchronous database insert.

Correctness and safety

Loss, overload, and schema drift are treated as observable states.

01

Bounded queue

The channel places a finite limit between request producers and the database worker.

02

Must-log policy

Important enforcement outcomes bypass ordinary sampling decisions.

03

Schema verification

Expected columns and encoded rows are checked as a contract rather than inferred at query time.

04

Timed flush

Low-volume periods still advance partial batches instead of waiting indefinitely for size alone.

Hyx’s observable advantage is joined attribution at the edge. The benchmark shows encoding is small on this host, but does not prove database capacity, retention economics, or query performance at fleet scale.

Interpretation

The evidence boundary is part of the result.

01

Serialization only

The 584.3ns figure excludes queue pressure, network transfer, server parsing, merges, disks, and ClickHouse queries.

02

Schema evolves

Sixty-six is the current repository width, not a timeless product constant. New context must preserve schema and consumer compatibility.

03

A row is not a trace

A joined edge record explains one request boundary; distributed origin work may still require trace propagation across services.

Evidence register

Every public claim keeps its provenance attached.

E1

Wide row

The current LogRow defines 66 fields across delivery, image handling, TLS/JA4, bot, WAF, signals, model, network, fleet, and sampling context.

E2

Ingestion contract

The client implements bounded channels, batch size, periodic flush, sampling, must-log rows, and schema verification.

E3

Focused suite

Sixteen tests passed with zero failures; two performance benchmarks remain ignored in normal test runs.

E4

Encoding benchmark

Release JSONEachRow encoding measured 584.3ns per row, or approximately 1.71 million rows per second, on this host.

Research archive

Return to the systems studies.

Mechanisms, measurements, and the limits required to interpret both.

Open the archive