Topic: architecture
-
Pipeline, fan-out, orchestrator and critic panel: which multi-agent pattern fits which task
A pipeline fits tasks with a fixed sequence of transformations; parallel fan-out fits independent sub-questions or repeated attempts to vote on; an orchestrator with workers fits tasks whose decomposition is only known at run time; a critic panel fits outputs that need review against several criteria; each multiplies token cost and adds a coordination layer that can fail on its own.
-
DACI and RACI for technical decisions: one approver, named contributors
DACI names a driver who runs the decision, a single approver who makes it, contributors who have a voice but no vote, and informed parties who learn the outcome; RACI assigns responsible, accountable, consulted and informed roles to tasks. Either works for technical decisions when the roles are written down before the discussion starts.
-
.NET dependency injection conventions: lifetimes, scopes and the captive-dependency trap
Microsoft.Extensions.DependencyInjection registers services on an IServiceCollection with transient, scoped or singleton lifetime and injects them through public constructors; the documented rules are never to inject a scoped service into a singleton, to let the container dispose what it created, to avoid service-locator calls, and to validate scopes so that captive dependencies fail at start-up instead of leaking state across requests.
-
Event sourcing and CQRS: what they buy and what they cost
Event sourcing stores every state change as an immutable event and derives current state by replay; CQRS separates the write model from read models. Both add auditability and flexibility at the price of complexity and eventual consistency.
-
Choosing Go or Rust for a new service: a decision procedure without benchmarks
Decide between Go and Rust for a service from documented language properties and team constraints rather than from benchmark folklore: memory management model, error and concurrency style, the shape of the workload, the libraries the service must talk to, and who will maintain it in two years.
-
Technical debt as a deliberate decision with bookkeeping
Cunningham's metaphor treats not-quite-right code as a loan, with every minute of extra work on it as interest. The metaphor only holds if the debt is taken on deliberately, recorded, and reviewed regularly; as a catch-all for anything ugly or an excuse for sloppiness, it is worthless.
-
Selecting a tool or skill with a decision model: Choice to rank, Noul to abstain
Why a Choice over candidate tools answers a relative question (which candidate fits best) while a Noul per candidate answers an absolute one (does this turn need a tool at all), and how TypeSafe's skill-suggestion cookbook combines both over a catalogue of 182 skills: one request ranks all, a second reads the top three and may reject all of them.
-
File upload service walk-through: direct-to-storage tickets, asynchronous scanning and quotas
A design walk-through for uploads that bypass the application servers: a ticket that reserves quota and returns a signed upload URL, a completion step that verifies the stored object, a scan worker that promotes or deletes it, lifecycle rules for abandoned uploads, and a status model that explains every stored object.
-
Notification service walk-through: channels, preferences, delivery attempts and retries
A design walk-through for a multi-channel notification service: an ingest API with an idempotency key, a router that expands one notification into per-channel deliveries after checking preferences, per-channel queues and retry schedules, callback handling for invalid tokens, and what to defer.
-
Storing derived data in PostgreSQL: generated columns versus materialized views
A generated column derives one value per row from that row alone, either virtual (computed on read) or stored (computed on write), and may use only immutable expressions; a materialized view stores the result of any query and is only as fresh as its last REFRESH. Use stored generated columns for per-row normalisation that should be indexed, materialized views for expensive aggregates that may lag, and a maintained summary table when neither fits.
-
Agent memory design: what to persist, what to summarise and what to forget
An agent's memory has three tiers: the context window, a task scratchpad and a durable store across sessions; decide per item which tier it belongs to, keep durable memory small and reviewable, and delete what is no longer true.
-
Server-sent events versus WebSockets
Server-sent events stream text events from server to client over plain HTTP with automatic reconnection and last-event IDs; WebSockets provide a bidirectional binary-capable channel with its own protocol. Choose SSE for one-way updates and WebSockets when the client must send frequently.
-
Abstaining as an agent: when not acting is the correct output
An agent's output space should include a deliberate 'not decided' for every automated action: what abstention is, why a classifier or agent without one converts every unclear case into a wrong action, and how to build abstention in through explicit options, floors on confidence, stakes-dependent thresholds and a route for what was abstained from.
-
When does client-side routing still pay off now that browsers offer bfcache, prerendering and cross-document view transitions?
Open question: in-page routers were adopted to avoid full page loads, at the cost of shell serving, 404 handling, scroll and focus restoration and a bundle that must arrive first; the back/forward cache, the Speculation Rules API and cross-document view transitions now address the original motivations in multi-page sites. For which sites and interaction patterns does an in-page router still measurably win?
-
At what repository size do teams need monorepo build tooling beyond plain Git?
Open question: sparse checkout, partial clone and per-directory CI filters cover the first stage of a growing single repository; at which size, team count or build time have teams found that a build graph tool with remote caching became necessary, and what did the transition cost?
-
Choosing between threads, processes and asyncio for a Python workload
Classify the hot path first: waiting on I/O suits asyncio (many connections, async libraries) or a thread pool (few blocking calls); pure-Python CPU work needs processes or a free-threaded build; native code that releases the GIL can use threads. Bound every pool, choose the process start method explicitly and write the shutdown path.
-
URL shortener walk-through: key generation, redirect status and abuse controls
A design walk-through for a URL shortener: random base62 keys with collision retry, a redirect path that touches one cache and one store, 302 rather than 301 when targets must stay revocable and countable, creation-side abuse checks, and a list of what not to build first.
-
Architecture decision records
An architecture decision record captures one significant decision with its context, the decision itself, its status and its consequences, in a short file kept with the code.
-
Application caching: expiry times, invalidation and stampede protection
An application cache holds derived data and may be lost at any time. It stays correct only through deliberate invalidation: an expiry time per data class, deleting instead of overwriting on writes, a version identifier in the key, and a lock against concurrent reloading of hot keys.
-
ETL versus ELT: where the transformation runs and what that changes
ETL transforms data in a separate engine before loading it into the target; ELT loads raw data first and transforms it with the target store's own processing. The choice decides where compute is paid for, what raw data lands in the warehouse, and how easily a transformation can be rerun.
Machine-readable: JSON