A0 · Overview — The Art of BCS
The Branded Component System taught the law; this course tests it in production. For a stateful, soft-real-time system run by one organization, the BEAM subsumes the four systems a production exchange is conventionally built around — a coordinator, a log broker, a message broker, and an orchestrator — and the branded identity carries the system’s state across every store and language.
The senior continuation of /bcs, taught from one working exchange platform. Every figure on every page is a committed gate transcript or a cited source. The claim is narrow on purpose: the regulated ledger, deep retention, and the observability sink stay outside the runtime by design.
0.999⁵ ≈ 99.50% — about 2.30 nines, narrowly
above the two-nines floor and far below the seven-nines target. The ceiling sits below the weakest single SLA,
and it falls as the stack grows.The BEAM — the virtual machine beneath Erlang and Elixir — was built at Ericsson from 1986 for very large numbers of concurrent, supervised, distributed, soft-real-time activities that must not stop. It provides, as language constructs, the properties the constellation supplies as infrastructure: lightweight processes for concurrency, supervision trees for self-healing, distributed Erlang for membership and discovery, bounded buffers for back-pressure, and hot code loading for change without downtime. The figure most often attached to that lineage is the AXD301 switch’s nine nines — which Armstrong’s own thesis reports with reservations, noting the long-term evidence was not collected systematically — so the course reads the nine-nines BEAM as a statement about design, not a guaranteed outcome.
The claim is narrow and measured, never anti-cloud zealotry: the patterns the course declines earn their keep at hyperscale and across many organizations. Three load-bearing claims carry it.
Claim one
Each member is a separate distributed system acquired to supply a property the runtime already has. You pay for it three times — in the dollars of running it, the latency of crossing it, and the availability it subtracts when it fails — to obtain what the BEAM provides for the cost of ordinary code.
paid in the currency that matters · availability
Claim two
Managed does not mean never-down, and an SLA is a credit schedule, not a guarantee. The arithmetic above is why, even in the months every piece meets its SLA, the composed product can sit below the line a real-time system needs.
and when the dependency is down · you wait
Claim three
A soft-real-time system meets its deadlines predictably; the constellation’s failures are stalls on the path — a quorum re-electing, a consumer group rebalancing, a log partition unavailable mid-write. Owning the engine removes the third-party terms whose product capped you.
achievable availability becomes a function of your own design
The line the course does not cross: the regulated, double-entry record of settled money is PostgreSQL; deep retention and disaster recovery are object storage; observability terminates in a hosted sink. Each stays outside the runtime on purpose. What the course retires is the coordinator, log, broker, and orchestrator constellation — not the pieces where a property truly belongs outside.
State the cloud’s case fairly first. The highest commitment AWS makes for compute is a
monthly uptime of 99.99% for instances spread across availability zones, with a single instance committed to
99.5%, and no service offered at 100%. Availability composes by multiplication, so put the published numbers
through aN — the same function the calculator above computes:
five at three nines 0.999 ^ 5 = 0.99501 ← 99.50% · below the weakest single SLA five at four nines 0.9999 ^ 5 = 0.99950 ← 99.95% · the cloud’s best, still under five nines the credo’s target seven nines ≈ 3 s / year ← unreachable while a third-party term remains downtime at 99.50% 0.00499 * 8760 h ≈ 43.7 h / year ← from the dependencies alone
The ceiling sits below the weakest single SLA, and it falls as the stack grows. This is the arithmetic under the conviction the course is written from: chain enough borrowed availability and the product drops regardless of how good each piece is. The failure is not only that the nines are too few — it is that the missing time arrives as a spike on the hot path, where a matching engine can least absorb it, injected by a system that cannot be inspected while it happens.
Each member of the constellation maps to a runtime primitive rather than an external service. Select one to read what it is for, and the runtime answer that subsumes it.
Coordinator
→ distributed Erlang
ZooKeeper’s leader election, membership, and configuration — held by distributed Erlang, the
single-writer principle, and the audited hash32 placement.
Log broker
→ Valkey streams
Kafka’s partitioned, replicated, retained log — below its envelope, per-instrument Valkey stream lanes, with three messaging shapes each in its own primitive.
Message broker
→ EchoMQ lanes
ActiveMQ / Artemis as a separate JVM daemon — the same work as a library on the database the platform already runs: fair lanes, at-least-once, pause, resume, depth.
Orchestrator
→ supervision + Fly + FLAME
Kubernetes’ scheduling, scaling, healing, discovery, rollout — supervision at the process, Fly machines as the grain, releases as rollout, FLAME for elastic compute.
What BCS adds — identity across boundaries
A runtime primitive coordinates within a runtime; a production system’s state lives in tables, on streams, in a Postgres ledger, in a cache, and in another language’s worker. The branded snowflake — a fourteen-byte identity, typed, time-ordered, self-placing, and identical across the canon — makes a fill matched by a process the same addressable entity in the ledger that records it, the stream that logs it, the cache that serves it, and the worker that prices it. The runtime gives the coordination; the brand gives the identity that makes the coordination span systems.
The credo and its arithmetic, the runtime, and the map — and the course’s design exemplar.
A1Part I — the constellation, the BEAM primitives that subsume it, and identity across boundaries.
A2Part II — ZooKeeper’s role on distributed Erlang, the single writer, and the audited hash.
A3Part III — Kafka below its envelope: per-instrument streams and the three messaging shapes.
A4Part IV — Artemis as a JVM daemon against EchoMQ as a library on the running database.
A5Part V — Kubernetes’ five jobs, each with a smaller answer on the BEAM, Fly, and FLAME.
A6Part VI — LMAX on the BEAM: the Disruptor seat, latency by construction, the matching core.
Part VII — the journal and shadow inside, the regulated ledger and retention outside, the line drawn.
Part VIII — the footprint totalled, the absent constellation, and the bridge to the deployment course.
[RECONCILE]
The foundation: the law in three clauses and the 14-byte branded-snowflake identity contract, taught from the manuscript and proven by frozen transcripts.
this course assumes the law, the contract, and the as-built ring, journal, and lanes — and doors here wherever it leans on them.
/echo-persistence[RECONCILE]
the near neighbor[RECONCILE]
/echomqThe Valkey-native bus and stream tier, rung by rung: the keyspace, the Lua inventory, the conformance suite.
this course doors here where the lanes, the claim-check bus, and the per-instrument stream log appear — A3 and A4.
/redis-patternsThe substrate patterns under the bus and the cache: sorted sets, atomic Lua moves, locks, streams — the judgement layer.
this course doors here for the Valkey moves the coordinator, log, and broker arguments rest on.
Sources
Related