Decisions
20
Decisions index · 20 of them, across 25 systems
This is the architectural record of twenty-five production systems: what was decided, why, and the price the decision carries. Ordered by the judgment it took rather than by date, so the hardest calls come first. Client work is under NDA, so the clients are unnamed and the decisions are not.
Decisions
20
Systems drawn from
13
With a cost stated
20
Under NDA
11
client unnamed, included anyway
01
Why
Conflict resolution is a business decision, not a technical one. The technology can detect the conflict; only the business can say who is right.
What it costs
The rule is domain-specific. It is correct for parcels and shop visits and would be wrong for a trading desk where a cancellation carries legal weight. Pretending it is universal would be the actual error. A rising escalation rate is the system being honest that reality contains a case the rules do not cover; the failure mode to fear is an escalation rate of zero, achieved by guessing.
02
Why
A system that silently overrides a human decision is a system that human will fight, work around, and eventually disable. An automated decision nobody understands is worse than no automated decision, because now there is both a wrong answer and nobody watching for it.
What it costs
Every resolution path has to produce prose, which is more work than returning an enum. The reason string is the product; the resolution is just the mechanism.
03
Why
Gaseous waste is scored under water pollution because it affects water. That is a domain distinction, not a modelling mistake, and flattening the three blocks into a uniform criteria set would destroy the meaning the instrument carries.
What it costs
Three asymmetric blocks are harder to aggregate across than one uniform set, the cost of keeping the domain intact.
04
Why
A conflict during reconciliation is information, not an error. Tagging it keeps the batch moving and puts the decision in front of a human who understands the business context.
What it costs
Someone has to review tagged conflicts, which is a workflow rather than a silent resolution, and that is the honest trade for a system handling money.
05
Why
A model producing well-formed, confident, conflicting output is the expected failure mode, not an edge case. Conflict detection runs as ordinary code against the result.
What it costs
The constraint logic exists twice, once as specification and once as code, duplication that buys the ability to reject a wrong answer.
06
Why
A pipeline whose input is a file and whose output is a file has no state to keep. Adding a database would create an operational surface for no benefit.
What it costs
No run history or resumability, the correct trade for a job measured in minutes.
07
Why
Rounding up can charge a customer more than the displayed price. Truncation is always in the customer favour and is defensible.
What it costs
A fraction of a unit is absorbed per transaction, which is the correct side to err on.
08
Why
A balance you can edit is a balance you cannot audit. Deriving it from an append-only history means any discrepancy has a traceable cause.
What it costs
More writes and a computed read, in exchange for being able to answer where a number came from.
09
Why
Only the three writes that must be atomic sit inside the session: freelancer credit, platform commission, deal status. Name lookups and notifications happen outside it.
What it costs
More code than wrapping the whole handler, but the transaction window stays small enough that lock contention never becomes the bottleneck under concurrent completions.
10
Why
Quantity changes in exactly one function, which writes a signed movement with the resulting balance. Nothing else touches quantity, so the audit trail cannot drift from reality.
What it costs
Every stock path routes through one function, a bottleneck by design and the reason the ledger is trustworthy.
11
Why
A ride copies the per-minute rate and unlock fee when it starts and never re-reads them, so changing the tariff cannot retroactively re-price a ride already in progress.
What it costs
Denormalised pricing on every ride, the right trade when the alternative is billing disputes.
12
Why
You cannot hold a database lock across nine hours of a driver being offline. The update carries the version the device last saw, and a zero-row result is the entire detection mechanism. One integer comparison, working identically whether the device was gone four seconds or four days.
What it costs
Conflicts are detected at sync time rather than prevented at write time, which is what makes arbitration necessary and is precisely the point. A stale write is not wrong, it is uninformed.
13
Why
Two riders scanning the same QR at once is the defining race in this domain. A conditional atomic update makes the guard and the lock the same operation, because the status filter is what prevents the second claim.
What it costs
No transaction means the physical unlock that follows needs an explicit compensating rollback. Without the status clause in the filter, both requests pass validation and two riders get billed for one scooter.
14
Why
The decrement filter carries the availability condition, so the database refuses the operation if stock is insufficient. There is no window between checking and writing.
What it costs
Error handling moves to interpreting a zero-match result rather than reading a friendly precondition first, which is less obvious to read and actually correct under concurrency. Without it, two tills selling the last box both read one remaining and both succeed.
15
Why
An offline client is an untrusted client. Without an allowlist, a modified device could write a revenue total or a price directly.
What it costs
New editable fields have to be added to the allowlist deliberately, which is friction that exists precisely so it cannot be skipped.
16
Why
Two systems never agree exactly on the current time. Without rewinding a short window, any product updated between the previous cursor position and the actual query silently never syncs.
What it costs
Some products are reprocessed each pass, which idempotent upserts absorb. That cost buys the guarantee that nothing is missed.
17
Why
One cluster, one deployment, one migration path. Onboarding a restaurant is a row rather than an infrastructure task.
What it costs
Every query has to be tenant-scoped, so that scoping lives in one data access layer where a missed filter is impossible rather than in every controller where it is inevitable.
18
Why
The product premise is that clients define their own content types. Storing field definitions as data is what allows onboarding without a deploy.
What it costs
Content fields cannot be indexed because names are not known ahead of time, and validation against declared types has to be rebuilt at the application layer rather than inherited from the database.
19
Why
The client business already runs on it. Replicating pricing and stock authority into a new system means two truths and a reconciliation problem nobody wants to own.
What it costs
The platform depends on the ERP being reachable, so read paths cache and write paths are designed to be replayable.
20
Why
Subscription and invoice are created before the gateway call, so a gateway failure has to unwind both rather than leaving an orphaned pending subscription.
What it costs
Two writes to undo instead of one transaction, the cost of an external system that has no concept of your transaction.
What this index leaves out
Fifty decisions are documented across the twenty-five systems. These twenty are the ones a stranger can judge me on. The rest sit on their project pages, including the four remaining decisions from the field operations sync engine, which are easier to read next to the conflict trace they came from.