Insights from real-world software architecture, system design, and engineering leadership
GraphQL let Facebook's clients write their own queries because no server team could predict every path through its graph. Every team importing it should show it has the same problem, and a team that locks GraphQL down to persisted queries has already listed every query it will send, so it needs GraphQL only when its screens take so many changing paths that an endpoint for each costs more than securing a query language.
Read post →Gating paid features with flag segments is easy to start and hard to undo, because which customers are owed a feature is a commercial decision that needs its own owner, audit trail, and offline rules. It belongs in its own domain, and a flag can hold it only as a deliberate stopgap with an owner and a written trigger for replacing it.
Read post →Business rules end up in the API gateway because it's the easiest place to apply a rule to every request, and there they lose their owner, miss traffic that skips the gateway, and leave the domain's tests. Keep a rule in the gateway only if the gateway team could change it correctly on its own, and admit a business rule only as emergency debt with an owner and a removal date.
Read post →API updates that send a wide body can't rely on clients to keep an omitted field distinct from a cleared one, whatever the HTTP method, and patch formats push that tracking onto every client. Normalizing the write surface avoids both, keeping reads as composed views and writes as small resources any client can replace whole.
Read post →Whether a service request is legitimate is treated as an authentication question, but the answer is also an architectural one: whether legitimacy comes from network position or from verified ownership. That distinction shapes service structure, security posture, and team dynamics in ways that compound as systems grow.
Read post →SOLID, normalization, least privilege, and bounded contexts share a structural belief: that where authority lives determines what the system can absorb. This post develops that into two properties for judging an authority claim, contour and bond, and traces through an order workflow how authority gets misplaced at the outset and how, once correctly placed, it erodes through locally reasonable optimizations.
Read post →Reporting pressure gradually distorts production schemas until they serve two masters and compromise for both. Separating the workloads lets each model evolve for the consumers it was designed to serve.
Read post →The web platform already does what most apps do, but Apple's control of the phone screen is the hardest constraint keeping the industry building native. The 'users prefer native' narrative is circular logic created by the constraint itself.
Read post →Many observability failures trace back to code that doesn't classify its own behavior. When your system can't distinguish 'handled correctly' from 'actually broken,' no platform can compensate.
Read post →Many software projects go wrong by drifting out of alignment with the need, the tradeoffs, and the risk. Realigning requires stopping, and since plan continuation bias keeps planning, discovery, coding, and testing from stopping, the fix is to build stops into the work instead of waiting to notice.
Read post →Shared libraries promise reuse and consistency but more often bind team autonomy and development tempo through coupling and coordination overhead. The consistency they claim to provide is better achieved by sharing principles, tradeoffs, and values rather than sharing implementation.
Read post →As AI makes code generation cheaper, the bottleneck shifts from technical execution to business judgment. The spec-executing middle is losing leverage, and this post argues careers will move toward either deep-and-mechanical or broad-and-human work.
Read post →Structure your data so invalid states cannot exist. Validate at construction, trust internally, and let null crash loudly rather than masking absence with defaults that propagate corruption silently through your system.
Read post →The 'SEO is dead' narrative draws a false binary between SEO and social media. Search authority now takes years to earn and favors incumbents over newcomers, but it still surfaces trusted sources, so use social channels for reach and speed and let SEO compound as a long-term credibility asset.
Read post →Sprints organize around time intervals. Shaped Kanban organizes around completing features with clear boundaries and circuit breakers to bound risk. Work flows at its natural pace within disciplined constraints.
Read post →The term 'tech debt' perpetuates the communication failures that created the problem. It's ambiguous, defensive, and invites deprioritization. Replace it with Corrections, Optimizations, and Re-Alignments to break the cycle.
Read post →Reproduction is the fulcrum of effective troubleshooting, because when the evidence fits more than one cause, without it you're guessing about the problem and guessing about the fix. The fastest path to a proven fix restores service first, then gathers facts, tests assumptions, and proves causation, and teams practice that discipline before the incident so it's there when they need it.
Read post →Many celebrated system rebuilds appear successful not because of new technology, but because they force teams to realign with value and best practices. This realignment work could often have happened without the rebuild.
Read post →Treating package updates as investments rather than chores means making deliberate, context-driven decisions based on value and risk instead of following dogma or chasing version uniformity.
Read post →Weighing the arguments for Result types and for exceptions when handling expected failures in distributed C# systems, and why the structural ones make Results the better default for domain operations.
Read post →Keeping secrets and environment-specific configuration alongside application code creates security risks and deployment coupling. A versioned, distributed config store solves these problems while introducing trade-offs worth making.
Read post →Personal development goals, family culture, and unlimited PTO sound empowering. Each can hide a leadership failure, and once a leader refuses an employee's request for the missing piece, it becomes a control mechanism that externalizes that failure through ambiguity, emotional manipulation, and blame-shifting.
Read post →A user's session token should be validated and consumed at the system's edge, then replaced with signed, request-bound internal context and service credentials. Forwarding the token inward couples internal services to the external auth mechanism, spreads a replayable credential, and breaks when work runs after the session or without one.
Read post →Architectural costs arrive late and tend to get blamed on the people working inside a design rather than the design itself. The fix is spending design time in proportion to how expensive a decision is to reverse, and pricing the migration rather than the code change.
Read post →Embedding authorization in JWTs creates security risks and UX problems because immutable tokens don't match dynamic permissions. Learn why looking up session-based grants at the edge is worth its small latency cost and stricter availability demands.
Read post →Some teams that say they use hexagonal architecture reach part of its goal, a database they can swap in tests, through layered code and framework dependency injection, without the structure Cockburn described. The difference shows up in who owns the interfaces and whether every driver can reach the business rules without a controller, and knowing which one you built tells you which benefits you actually have.
Read post →Why REST's resource-centric design conflicts with domain-driven architectures and how RPC provides better alignment with business operations.
Read post →Single-minded focus atrophies everything else you know. Real focus means being fully present with what matters right now, shifting attention as the work reveals what needs it.
Read post →Once you know a topic's basics, building real projects teaches more than passive learning platforms. Real capability comes from doing, failing, and fixing, not collecting completion checkmarks.
Read post →Used as a default gate, timed algorithm puzzles reward drilled pattern recall in roles that rarely use it. Structured exercises built from the role's own work predict about as well and test what the job needs.
Read post →Once code is correct and fast enough, what makes it good over time is how cheaply it absorbs change. Cleverness and premature optimization raise that cost, and flexibility pays only where change is likely.
Read post →TDD's real value isn't code coverage. It's catching wrong assumptions before you deliver the wrong thing.
Read post →Why monoliths are effective for discovery and microservices are optimizations: principles for choosing the right architecture for your context.
Read post →Leadership in software development comes from habits, not titles: accumulating experience instead of repeating it, holding your work to your own standard after it ships and when it passes review, looking for the questions you haven't asked, and growing others as you grow.
Read post →