HYX

Study 01 / cache systems

Hyx research study

Mobius under pressure.

A production-oriented study of scan resistance, bounded concurrency, and the measured performance gap between Hyx Mobius and TinyUFO.

Research classification

Internal measurements / repository invariants / public literature

E1 / finding
~1.2×
read throughput
Mobius relative to TinyUFO · controlled core-pinned measurement
E1 / finding
~2.7×
put throughput
Mobius relative to TinyUFO · low-concurrency measurement
E3 / finding
50K
scan insertions
hot object retained under sustained one-hit traffic
E4 / finding
800K
mixed operations
eight-thread bounded concurrency test

Abstract

The fastest policy is the one the hot path barely notices.

Research question / RQ-01

Can a web-cache policy preserve useful objects under hostile churn while making coordination and bookkeeping nearly disappear from the hot path?

01 / Context

Dong et al. (2025) introduced Mobius as a lock-free cache-eviction design for throughput-sensitive systems. Hyx uses a production adaptation as the in-memory engine behind its response cache and other latency-sensitive subsystems.

02 / Measured answer

1.2×

read throughput

2.7×

low-concurrency puts

Meaningful on reads. Dramatic on the mutation path. TinyUFO is the measured baseline.

03 / Publication boundary

This study covers observable behavior, contractual semantics, controlled measurements, and production constraints. It excludes internal mechanics that could be reconstructed from the description.

Evidence over implementation

Research lineage

The original paper defines Mobius. Hyx evaluates a production adaptation.

Published study / Dong et al. 2025

Lock-free eviction as a throughput problem.

Dong et al. (2025) present Mobius as a lock-free design derived from SIEVE and evaluate it with synthetic workloads and four large real-world traces.

The authors report 1.2×–8.5× concurrent-throughput improvements over state-of-the-art comparison policies while maintaining comparable hit ratios. Their RocksDB experiments report average gains of 4% for point reads and 14% for range reads over the LRU-backed system (Dong et al., 2025).

The paper also reports that Mobius and SIEVE share identical single-thread hit ratios under its synthetic evaluation, while Mobius improves throughput and tail latency through its concurrency design.

Read the original paper ↗

Hyx study / production evidence

Same lineage. Independently measured implementation.

Hyx’s engine follows the paper’s throughput-first research direction but is not presented as a byte-for-byte reproduction. Repository evidence records that a faithful version of the paper’s second optimization was implemented and measured before the production design intentionally diverged.

This public study reports only externally meaningful behavior: relative throughput, bounded capacity, scan resistance, TTL semantics, update guarantees, and concurrency invariants. It withholds the mechanics and constants needed to reconstruct the production path.

Baseline warning: Dong et al. (2025) use a different comparison system. Hyx’s ~1.2× read and ~2.7× write figures compare against its previous cache. The results use different baselines and are not interchangeable.

Evidence / E1 · E8

Comparative result

Mobius leads TinyUFO most sharply where caches change.

R-01 / controlled comparison

higher is better / relative throughput

Largest measured lead

2.7×

low-concurrency puts / TinyUFO baseline

The mutation path is where Mobius most clearly separates itself in the recorded comparison.

DimensionMobiusTinyUFOEvidence boundary
Read path~1.2× relative throughputBaselineControlled, core-pinned internal microbenchmark
Put path~2.7× relative throughputBaselineLow-concurrency internal microbenchmark
Update semanticsExisting value is deterministically replacedAdmission can reject a candidateProduction rules-cache contract
Pressure modelBounded by entry count or explicit weightCapacity-boundedRepository invariant tests

E1 / internal controlled microbenchmark · relative ratios are repository-recorded · not an independent third-party reproduction · not end-to-end CDN throughput

Repository experiments

Speed is interesting. Staying correct under pressure is decisive.

01E3

Adversarial scan

A resident hot object is repeatedly accessed while 49,999 distinct newcomers pass through a cache sized far below the stream.

Result / The hot object remains resident. This establishes scan resistance for the tested workload, not universal optimality.

02E4

Concurrent mutation

Eight workers perform 800,000 mixed reads, writes, and removals across a keyspace five times larger than the configured capacity.

Result / The test completes without worker panic and the reported cache weight remains within its configured bound.

03E5

Weighted pressure

A large newcomer enters a cache already filled with many lightweight objects; separate tests also restore entries at larger weights.

Result / The newcomer is retained while total weight remains bounded, covering non-uniform object sizes.

04E6

Sustained soak

Sixteen workers run mixed reads, writes, removals, and short TTLs against a keyspace twice the capacity for five seconds.

Result / The ignored release-mode soak asserts bounded weight and panic-free completion. It is a stress invariant, not a latency benchmark.

Production relevance

A cache engine is valuable when the surrounding system can trust its behavior.

O-01 / integration surface

one engine / multiple hot paths

01

Response delivery

Fresh, stale, variant-aware, range-capable response serving with a disk reserve beneath memory.

02

Compiled policy

Capacity-bounded caches for domain rules and negative lookups, with deterministic replacement on updates.

03

Optimized assets

Weighted storage for transformed image outputs where object size, rather than object count, defines pressure.

04

Revalidation

Stale-while-revalidate and request coalescing reduce duplicate origin work around cache misses.

Interpretation

What these results establish, and where they stop.

01

Internal measurement

The TinyUFO ratios are controlled measurements recorded in the Hyx source. They are not presented as independent third-party results.

02

Host-specific absolute cost

Nanosecond-scale cache measurements vary with CPU, allocator, compiler, pinning, object type, and contention. Relative results are more portable than absolute timings.

03

No universal workload winner

The tests cover selected scans, mixed mutation, weights, TTLs, and integration semantics. They do not prove optimal hit ratio for every trace or object distribution.

The results support a specific conclusion. Under Hyx’s production-shaped test requirements, Mobius preserves the required cache invariants and materially outperforms the previous TinyUFO engine, with the largest improvement on writes.

Evidence register

Every Hyx claim has a repository anchor.

E1

Controlled comparison record

The Mobius engine’s repository documentation records core-pinned measurements of ~1.2× TinyUFO reads and ~2.7× TinyUFO low-concurrency puts. Measurements use the deployed allocator. Raw host output is not published on this page.

E2

Production integration record

The response cache identifies Mobius as the sole in-memory engine after concurrency soak validation. The compiled-rules cache documents deterministic replacement behavior and hot lookups.

E3

Scan-resistance test

The repository test keeps one hot object alive through a 50,000-object scan against a 2,000-entry capacity.

E4

Concurrent stress test

Eight threads each execute 100,000 mixed operations; the final assertion requires measured weight to remain at or below capacity.

E5

Weighted-capacity suite

Heavy-newcomer, restore-at-larger-weight, count-mode, and flood tests assert that capacity accounting remains bounded.

E6

Mixed-workload soak

A release-mode, 16-thread soak combines TTL and mutation pressure and checks boundedness throughout and at completion.

E7

Cache semantics suite

Separate tests cover hit, miss, expiry, stale reads, removal, borrowed-key lookup, duplicate prevention, and queue compaction bounds.

E8

Research-divergence record

The engine documentation records a measured faithful implementation of the paper’s consecutive-detection design and the evidence-led decision to use a faster production adaptation. Mechanics are intentionally omitted here.

Research archive

Return to the systems studies.

Read the evidence before the headline. That is the standard for every study Hyx publishes.

Open the archive