03 / NOTES & OBSERVATIONS
Thinking out loud.
In writing.
Long-form engineering essays, worked examples and readings of research papers. On systems, their edges, and the decisions that survive contact with the real world.
Notebook dates are assigned across June 2025–September 2026 to organise this collection. Each essay lists its publication date separately.
78 of 78 essays
SYSTEMS
14 ESSAYSEngineering essay · 7 MIN READ
Fuzzing the parser boundary
Build a parser fuzz target around bounded resources, explicit rejection and semantic oracles, then preserve useful counterexamples.
Engineering essay · 7 MIN READ
Floating-point is an error budget
Choose numeric representations, tolerances and stable algorithms from the meaning of the result rather than a blanket rule about decimals.
Engineering essay · 7 MIN READ
Arena allocation and the cost of reset
Arena allocation simplifies allocation only when object lifetimes, cleanup and retention agree with the reset boundary.
Engineering essay · 7 MIN READ
io_uring and buffer ownership
Model submitted buffers, terminal completions, cancellation and zero-copy notifications as an explicit ownership protocol.
Engineering essay · 7 MIN READ
Why “unknown” should be a valid system output
Abstention as a decision with a cost, and a Rust contract that keeps unsupported conclusions separate from negative answers.
Engineering essay · 7 MIN READ
Lock-free is not wait-free
Separate system-wide progress, individual completion and elapsed deadlines when evaluating concurrent algorithms.
Engineering essay · 7 MIN READ
Bounded queues are a product decision
Queue capacity determines which promises a service makes under load. Backpressure, rejection and deadlines have different user-visible meanings.
Engineering essay · 7 MIN READ
Property tests need a model
Generated inputs become valuable when an independent model, meaningful invariants and intentional generators define what should happen.
Engineering essay · 7 MIN READ
Rust async without a runtime: what that actually means
Futures define a polling contract. Executors, wakeups, I/O drivers and scheduling policy are separate responsibilities that still need an owner.
Engineering essay · 7 MIN READ
When Go is the right rewrite
A rewrite should remove a measured constraint while preserving the service's real contract. Go's runtime is a tradeoff to evaluate, not a verdict.
Engineering essay · 7 MIN READ
Cancellation is a protocol
Stopping local waiting does not undo remote work. A reliable cancellation design needs ownership, acknowledgements and a clear commit boundary.
Engineering essay · 7 MIN READ
State machines need invalid transitions
A useful state machine specifies rejected events, guards, atomic transitions and recovery as carefully as its happy path.
Engineering essay · 7 MIN READ
Designing systems for uncertainty
Freshness, provenance and conflicting observations belong in the data contract, before a consumer decides what to do.
Engineering essay · 7 MIN READ
Memory ordering without mysticism
Reason about atomic publication through happens-before edges, ownership and concrete failure cases rather than processor folklore.
PERFORMANCE
20 ESSAYSEngineering essay · 7 MIN READ
Latency budgets need critical paths
Budget user-visible latency through dependency graphs, queue boundaries and deadlines rather than adding every span or percentile.
Engineering essay · 7 MIN READ
False sharing is a layout bug
Distinguish logical independence from coherence traffic, then change layout only when ownership and measurements support it.
Engineering essay · 7 MIN READ
Zero-copy is a lifetime decision
Avoiding a copy changes who owns memory, how long buffers stay alive and which boundaries can safely share them.
Engineering essay · 7 MIN READ
Indexes are a write-path decision
Evaluating indexes as maintained data structures with consequences for updates, storage, concurrency and the full query workload.
Engineering essay · 8 MIN READ
Small models and the cost of a decision
Reason about tiny learned runtimes through total decision latency, feature costs, precision, fallback behavior and honest benchmarks.
Paper analysis · 7 MIN READ
The tail is the system
Reading Dean and Barroso through fan-out, correlated slowdowns and the resource budget that makes hedging succeed or fail.
Engineering essay · 7 MIN READ
CPU time is not wall time
Separate execution, waiting and parallel work before turning elapsed time or CPU utilization into a performance diagnosis.
Engineering essay · 7 MIN READ
Cache locality before cleverness
The representation that exposes useful bytes can matter more than a clever loop. Start with access patterns, working sets and ownership between cores.
Engineering essay · 7 MIN READ
Profile the production shape
A profile estimates where a particular workload spent resources. Make its population, sampling and attribution boundaries explicit.
Engineering essay · 7 MIN READ
SIMD begins with a data contract
Vectorization depends on aliasing, numeric semantics, tails and deployment capabilities before instruction width becomes useful.
Engineering essay · 7 MIN READ
Prefix caching needs an identity
A prefix cache is correct only when its key identifies the computation, its authorization scope and the inputs that produced the cached state.
Engineering essay · 7 MIN READ
Performance regressions need experiments
Design comparative CI experiments around pairing, practical effect sizes and controlled environments instead of brittle timing thresholds.
Paper analysis · 7 MIN READ
The KV cache is a systems problem
A PagedAttention reading focused on memory accounting, fragmentation, ownership, scheduling and serving trade-offs.
Engineering essay · 7 MIN READ
7.4 microseconds: what a benchmark must explain
A reported median is the beginning of a performance argument. Measurement boundaries, sampling and workload design determine what it means.
Engineering essay · 7 MIN READ
Open-loop load tests tell a different story
Arrival schedules, coordinated omission and generator limits determine what a load test can honestly say about overload.
Paper analysis · 7 MIN READ
Mixture of experts still needs a scheduler
Sparse expert activation saves arithmetic only within a system that can route tokens, balance capacity and control communication costs.
Engineering essay · 7 MIN READ
Bloom filters spend false positives
Sizing Bloom filters, pricing false positives and preserving the conditions under which a negative answer is trustworthy.
Engineering essay · 7 MIN READ
Syscalls are not the whole I/O cost
Account for copies, cache state, queueing and durability before attributing I/O performance to system-call counts.
Engineering essay · 7 MIN READ
LSM trees move the write cost
Understanding log-structured storage through compaction debt, read and space amplification, tombstones and sustainable throughput.
Engineering essay · 7 MIN READ
Quantization moves the error
A worked quantization and bias-correction example shows why small weight error does not guarantee small decision error.
AI
21 ESSAYSEngineering essay · 7 MIN READ
Context compression can delete the proof
Token reduction is useful only when compressed context preserves the evidence, exceptions and uncertainty needed to justify the answer.
Engineering essay · 8 MIN READ
Backpressure for streaming AI
Bound token and audio streams with explicit queue budgets, cancellation, fair scheduling and honest completion semantics.
Engineering essay · 7 MIN READ
Tool permissions need capabilities
Agent tools need narrow, enforceable authority over operations and resources, with explicit limits for delegation, expiry and commitment.
Engineering essay · 7 MIN READ
Embeddings are a versioned interface
An embedding migration changes the meaning of distance. Treat the encoder, corpus and index as one versioned retrieval contract.
Engineering essay · 7 MIN READ
Evaluating AI without grading the vibes
Build evaluations around decisions, paired comparisons, error costs and reproducible conditions rather than polished answers.
Paper analysis · 7 MIN READ
Speculative decoding spends parallelism
The exact sampling correction behind speculative decoding, with a worked acceptance example and realistic cost boundaries.
Engineering essay · 7 MIN READ
Synthetic data needs a failure model
Generated examples can expand coverage or erase it. Analyse selection bias, recursive replacement and verification before treating volume as new information.
Engineering essay · 7 MIN READ
Prompt injection is a data boundary
Defensive design for keeping retrieved text, tool output and model suggestions from acquiring application authority.
Engineering essay · 7 MIN READ
AI evaluation needs a dataset lifecycle
A benchmark is a versioned measurement instrument. Track provenance, split related cases together and make corrections without silently rewriting the result.
Engineering essay · 7 MIN READ
Chunking is an information boundary
A retrieval chunk decides which qualifiers, relationships and source identities remain available when a passage is read alone.
Engineering essay · 7 MIN READ
Voice AI should prepare before you stop talking
A design for speculative voice responses that separates listening, preparation, commitment and cancellation.
Engineering essay · 7 MIN READ
Speech interruption is a state machine
Stopping a voice assistant requires more than cancelling generation. Invalidate stale audio, control playback and reconcile what the user actually heard.
Paper analysis · 7 MIN READ
LoRA is a capacity choice
A reading of LoRA through rank, parameter counts, deployment identity and the limitations of low-rank adaptation.
Paper analysis · 7 MIN READ
Attention is a memory interface
Reading Attention Is All You Need through addressing, data movement, masking and the limits of weighted retrieval.
Engineering essay · 7 MIN READ
Model routing needs a fallback contract
Routing between models needs explicit quality, latency, privacy and action semantics when the preferred path fails or escalates.
Engineering essay · 7 MIN READ
Retrieval needs an abstention policy
Why retrieved text is not enough: evidence contracts, claim-level support and risk-versus-coverage decisions for RAG.
Engineering essay · 7 MIN READ
Reranking is a budget allocation problem
How candidate recall, cross-encoder cost and evidence coverage determine where a retrieval system should spend its reranking budget.
Engineering essay · 7 MIN READ
Decoding is part of the product
Temperature, truncation and stopping rules change the behaviour users receive. Treat decoding as a versioned product policy, with task-specific evaluation.
Engineering essay · 8 MIN READ
Agents need transaction boundaries
Design AI tool execution around proposals, durable intent, idempotency, stale state and uncertain external outcomes.
Engineering essay · 8 MIN READ
Structured output is not semantic correctness
A typed AI response can still be false, stale or unauthorized. Design validation around the decision it will drive.
Engineering essay · 7 MIN READ
Agent memory needs forgetting
Durable agent memory needs retention, provenance, correction and deletion semantics that remain valid across summaries and indexes.
INFRASTRUCTURE
10 ESSAYSEngineering essay · 7 MIN READ
Disaster recovery is a replay budget
Deriving recovery objectives from durable copies, restore throughput, replay capacity, dependency recovery and verified cutover.
Engineering essay · 7 MIN READ
Cross-chain funding attribution: evidence before interpretation
A rigorous model for cross-chain traces that separates observations, bridge joins, conservation checks and uncertain attribution.
Engineering essay · 7 MIN READ
Tracing needs context, not just spans
Building useful traces through causal propagation, asynchronous boundaries, sampling and controlled diagnostic context.
Engineering essay · 7 MIN READ
Retry budgets prevent amplification
Treating retries as additional load with a finite budget, coordinated deadlines and explicit rules for ambiguous operations.
Engineering essay · 7 MIN READ
Cardinality is an observability budget
Budgeting metric series through label combinations, histogram expansion, churn and the questions each dimension must answer.
Engineering essay · 7 MIN READ
Rate limits need a fairness policy
Why a request counter cannot decide fair access, and how identities, cost weights, concurrency and reserved capacity change admission.
Engineering essay · 7 MIN READ
Backfills that do not fight production
Design resumable backfills around consistent cuts, conditional writes, adaptive resource budgets and verifiable completion.
Engineering essay · 7 MIN READ
Zero-downtime migrations need expand-contract
Planning database changes as overlapping application versions, bounded data work and evidence-based removal of old schema.
Engineering essay · 7 MIN READ
Replay is an architectural feature
A retained event log becomes a recovery tool only when inputs, transformations, side effects, checkpoints and output comparison support replay.
Engineering essay · 7 MIN READ
Pagination needs a consistency contract
Choosing between live traversal and stable snapshots, with deterministic keyset ordering and explicit cursor semantics.
DISTRIBUTED SYSTEMS
13 ESSAYSEngineering essay · 7 MIN READ
WebSocket reconnection is state reconciliation
Reconnect a live application with explicit cursors, consistent snapshots, bounded replay and application acknowledgements rather than a socket retry alone.
Engineering essay · 7 MIN READ
Cache invalidation is a versioning problem
Making cache freshness explicit with complete keys, source versions, guarded fills and recovery rules for missed invalidations.
Paper analysis · 7 MIN READ
Event time is not wall-clock time: reading the Dataflow paper
An analysis of the Dataflow model through late events, watermarks, revisions and the application decisions that a streaming engine cannot make.
Engineering essay · 7 MIN READ
The dual-write problem: what an outbox actually guarantees
An outbox makes a database change and publication intent atomic; delivery, ordering, deduplication and consumer effects still need their own contracts.
Engineering essay · 7 MIN READ
Multi-region is a failure model
Designing regional resilience through quorum placement, replication lag, failover authority, capacity headroom and shared dependencies.
Paper analysis · 7 MIN READ
Raft is a replicated state machine contract
Reading Raft through its application boundary: commitment, deterministic execution, retries, membership changes and trustworthy reads.
Engineering essay · 7 MIN READ
Leases need fencing tokens
Why an expired lease cannot stop an old worker, and how resource-enforced fencing turns ownership changes into safe writes.
Engineering essay · 7 MIN READ
Isolation levels are observable behaviour
Understanding transaction isolation through concurrent histories, write skew, retries and the invariants users can observe.
Engineering essay · 7 MIN READ
Webhooks are a distributed systems problem
Treat webhook receipt, processing, ordering and reconciliation as separate contracts instead of assuming an HTTP callback is a reliable queue.
Engineering essay · 7 MIN READ
Configuration rollouts are distributed writes
Treating configuration activation as versioned state replication with validation, compatibility, canaries and recovery from missed updates.
Paper analysis · 7 MIN READ
Ordering events without a global clock: reading Lamport
Lamport clocks preserve causal order without measuring elapsed time; this analysis works through their guarantees, limits and practical alternatives.
Engineering essay · 7 MIN READ
Exactly once needs a boundary
Exactly-once claims become useful when they specify the effect, transaction boundary, identity, retention window and recovery assumptions.
Engineering essay · 7 MIN READ
Schema evolution without flag days
Evolve distributed contracts with explicit compatibility directions, semantic migration, expand-and-contract rollout and archived replay fixtures.
Paper analyses credit the original work; the arguments and worked examples are my interpretation. Subscribe via RSS for new essays.