Newest first, grouped by year. Search or filter by topic.
The series retrospective - what fourteen parts of reading Pacco as a system of systems actually taught, the absences the estate confesses when you list them together, how the successor evolved past DShop and where it stopped, and what the MIT licence makes of all of it.
The estate's full who-publishes-what matrix, built from the source - and the one broken cell in it. Orders' copied ParcelDeleted event declares the deliveries exchange while Parcels publishes on parcels, so Orders never learns a parcel was deleted, and neither the Pact tests nor the hand-maintained registry can catch it.
When a Chronicle saga rejects, the saga log is replayed backwards and every handled message gets its CompensateAsync called. Tracing both rejection paths through the source reveals a surprise - only one of them actually compensates - plus a ForEach that never awaits.
Pacco's naming system is consistent enough to parse - exchanges are bounded contexts, queue names spell their own bindings, commands and events conjugate in snake_case, and every command has a _rejected twin. Reading the estate as a language, including its irregular verbs and the fact that no grammar-checker exists.
Pacco traces every request, scrapes every service every five seconds and ships every log line to Seq - then forgets all of it on docker-compose down, because the telemetry stores are the containers whose volume blocks are commented out. The wiring is exemplary; the memory is the finding.
Pacco's umbrella repository contains no code, yet it decides everything about how the estate actually runs - seven compose files in deliberate layers, two PM2 manifests that quietly forget the saga service, a 460-line docker run-book, and a .rest file that is the best onboarding document in the system.
Pacco is thirteen sibling repositories held together by a solution file of relative paths, clone scripts with a dual-branch habit, and 284 package references that all say 0.4.* - coherence by wildcard, with a name-squatted saga library and a README that forgot a repo along the way.
Pacco's Identity service fits its whole auth machinery - JWT issuance, a refresh-token aggregate, Redis deny-list revocation - in about fifteen files, and most of them are right. The three that aren't teach more - a mapper bug that leaves 202 callers polling forever, a dead password check, and a refresh token that never rotates.
Pacco names every control - JWT validation, a Vault, a log scrub-list, admin-only routes - and then a sign-up body chooses its own role, the signing key lives in every repo, Vault is off in Docker and anonymous callers pass every ownership check. A governance read of which shortcuts survive copy-paste.
One JSON blob born at Pacco's gateway rides the Correlation-Context HTTP header into services and the message_context AMQP header across the bus, surfacing in SignalR pushes - carried by a class copy-pasted into eleven repos and a broker that forwards exactly one header.
Pacco's Operations service narrates the estate twice - once over SignalR with a JWT handshake and per-user groups, once over a gRPC stream with no auth, no filtering, a per-call event-handler leak and a blocking Take. Same data, same service, two security models.
The provider side of Pactify - how CreateFor<Startup> hosts your whole API in memory with TestServer, why interactions verify concurrently, the SmartFormat trick behind endpoint templating, and the discovery that Given(...) never runs.
Pacco's async gateway answers a POST with two names - the ID a resource will have and the operation that tracks whether it ever gets it. Following both through the Redis state machine, the saga header override and the per-user SignalR push, warts and all - a five-minute memory, a 404 with three meanings, and a status API anyone can read.
Pacco ships a second API gateway built on Ocelot that hand-rolls everything the Ntrada YAML declares - auth bypass, async bus routes, edge-minted IDs, correlation headers - plus the estate's only circuit breakers. It is instructive, revealing, and deployed by absolutely nothing.
FeedR pushes live prices to clients through a gRPC server stream: a four-message proto contract, fixed-point decimals, dual Kestrel endpoints, and an event-to-stream bridge that busy-spins - plus the channel-based rewrite it's asking for. Part 5 of the FeedR deep dive.
Pacco's Ntrada gateway is ninety lines of C# and five hundred of YAML, and swapping one route key - use downstream for use rabbitmq - turns the same HTTP surface from a proxy into a bus publisher answering 202. The two compose files disagree about which one you get.
Convey.WebApi threw away controllers, replaced every MVC formatter with its own, and bound route values into immutable commands by writing to compiler-generated backing fields. A source read of a routing layer built years before .NET shipped the same idea.
Pacco's Operations service listens to all eighty messages in the estate while owning zero contract classes - it reads a JSON registry and forges the CLR types at runtime with Reflection.Emit. The estate's answer to the shared-contracts problem, and the drift it cannot see.
Thirteen repositories, one umbrella that runs twelve of them, two interchangeable API gateways and a service that watches everything - the Pacco estate read as a system of systems, and the fifteen-part series it earns.
Seeker, initializer, processor, post-processor - Chronicle's whole runtime is four internal classes you can read in ten minutes. A source walk through the pipeline, including the gate that drops messages, the finally block that always persists, and the state that isn't as final as it looks.
The series retrospective - Pacco's domain teaches real rules (state machines, priority expropriation, loyalty ladders, compensations) on top of cardboard data (bare-decimal prices, one vehicle per order, size-cube parcels, no payment, and a capacity check that ships in two halves nobody ever connects). What to steal, what to distrust, and what fourteen parts of source-reading add up to.
AssignVehicleToOrder is the one command in Pacco that cannot finish alone - Orders calls Vehicles, then Pricing, and Pricing calls Customers, a synchronous enrichment chain on the write path of a messaging estate. At the end of it sits a deliberately bus-less calculator with a loyalty ladder, an unassigned Id and a discount logged in dollars.
Availability guards every save with a version-filtered ReplaceOneAsync - and discards the result, so a lost update is indistinguishable from a successful one and the integration event ships either way. Three folders away, Orders and Customers inherit the same AggregateRoot name with the versioning quietly deleted.
In Pacco the CQRS split is not a diagram - it is wiring you can read. Every command arrives through two front doors, RabbitMQ subscription and controller-less HTTP dispatch, while queries ride HTTP only; commands assign their own ids in the constructor, and a marker attribute turns each service into its own runtime contract registry.
Every Pacco service that enables Convey's Mongo outbox also sets disableTransactions to true - the aggregate write and the outbox write are separate operations, the pattern's one guarantee switched off in configuration. Meanwhile the inbox that would absorb the resulting duplicates sits beside three handlers that throw on them.
