Abstract
The claim begins with a question, not a headline.
Research question / RQ-06
Can a tiered CDN cache combine aggressive reuse with the authorization, validation, variation, and partial-response constraints of HTTP?
Mobius answers how memory is retained under pressure. This study asks a different question: whether the response cache preserves the HTTP contracts that decide what may be stored, reused, revalidated, ranged, or served stale.
Method
Audit cacheability, freshness, validators, variants, stale controls, coalescing, ranges, metadata encoding, disk promotion, and capacity tests against HTTP caching requirements.
Finding
The cache is structured around response semantics and concurrent state transitions. Eviction is only one bounded subproblem inside a larger protocol machine.
Disclosure boundary
The study publishes externally observable contracts and test categories, not proprietary key composition, production thresholds, or internal coordination mechanics.
System boundary
A cached object carries policy, representation, and time.
S-01 / system topology
admit / vary / age / validate / serve
Admission
Method, status, Authorization, Cache-Control, Set-Cookie, and policy determine whether a response is reusable and under which constraints.
Representation
Vary-derived variants, typed response heads, validators, encoding, and byte ranges prevent one stored body from being served under the wrong request contract.
Freshness
Age, explicit lifetimes, heuristic rules, stale-while-revalidate, and stale-if-error govern reuse over time.
Coordination
Per-key revalidation locks and miss coalescing keep concurrent misses from multiplying origin work while memory and disk tiers remain bounded.
Controlled measurement
The measured nanoseconds belong to two narrow parsing gates.
M-01 / host-local evidence
release mode / scoped result
| Measurement | Observed | Scope |
|---|---|---|
| 01No Range header gate | 2.49ns | early path before range parsing |
| 02Single range parse | 16.24ns | synthetic header benchmark |
| 03Focused suite | 125 passed | 3 benchmark tests ignored by default |
| 04Storage model | memory + disk | bounded reserve and promotion semantics |
Release-mode microbenchmarks on this host. These are not hit latency, origin latency, or end-to-end CDN throughput.
Findings
Correct reuse is a sequence of proofs.
A key is not enough
Vary and authorization rules determine whether two apparently identical URLs may share a stored representation.
Stale is policy
Stale-while-revalidate and stale-if-error are bounded permissions, not a generic license to return old bytes.
Validation mutates metadata
A 304 response merges selected metadata into the existing representation without replacing the stored body.
Concurrency changes semantics
Coalescing and per-key revalidation ensure one logical refresh does not become an origin stampede under load.
Correctness and safety
Protocol edge cases are encoded as executable constraints.
Authorization boundary
Shared reuse of authorized responses requires the explicit permissions defined by HTTP caching semantics.
Corruption rejection
Versioned cached heads reject incompatible or malformed metadata rather than constructing a response from ambiguous bytes.
Range discipline
Ranges are served only when the stored representation, validator, and requested interval permit a correct partial response.
Bounded tiers
Memory weight and disk budget are enforced independently; promotion cannot make capacity unbounded.
The repository evidence supports broad HTTP cache semantics and extensive focused tests. It does not establish a universal hit ratio, origin offload percentage, or end-to-end latency advantage because those depend on traffic and deployment.
Interpretation
The evidence boundary is part of the result.
Not an eviction sequel
This study does not repeat the Mobius/TinyUFO policy comparison. It examines protocol correctness around the stored object.
No traffic trace
Without a representative request trace, the study cannot claim a fleet hit ratio or stale-service frequency.
Microbenchmark boundary
The nanosecond figures cover range gates only, not metadata lookup, body transfer, disk IO, or origin revalidation.
Evidence register
Every public claim keeps its provenance attached.
Semantic suite
125 tests passed across RFC-oriented behavior, response metadata, variants, ranges, coalescing, negative caching, disk reserve, and concurrency.
Typed metadata
Cached response heads preserve multi-value headers and validators while rejecting incompatible versions and malformed data.
Concurrent refresh
Tests constrain stale-while-revalidate locks, miss coalescing, 304 merge behavior, and bounded tier promotion.
Range microbenchmarks
The host measured 2.49ns for the no-range gate and 16.24ns for the range parser in release mode.
References / Harvard style
HTTP caching
Fielding, R., Nottingham, M. and Reschke, J. (2022) HTTP Caching. RFC 9111. Internet Engineering Task Force.
Open source ↗Reference / 02HTTP semantics
Fielding, R., Nottingham, M. and Reschke, J. (2022) HTTP Semantics. RFC 9110. Internet Engineering Task Force.
Open source ↗Reference / 03Stale controls
Nottingham, M. (2010) HTTP Cache-Control Extensions for Stale Content. RFC 5861. Internet Engineering Task Force.
Open source ↗Research archive
Return to the systems studies.
Mechanisms, measurements, and the limits required to interpret both.