Version the prompt and tools with the model
A model identifier alone cannot reproduce an application response. Capture the configuration that shapes its inputs, available actions and accepted outputs.
Read articleCobnex blogs / Global
Cobnex blog
AI implementation, software architecture and cloud operations for teams worldwide.
456 articles
Page 4 of 26
Draw the request path before choosing the region
Start with every service that receives task data. A regional model endpoint does not describe the location or access rules of the rest of the pipeline.
Read articleAttach one budget to the whole task
Give every model call and tool attempt the same task identity, then reserve capacity before dispatch. This makes cost controls work across retries and concurrent workers.
Read articleDesign three outcomes for missing evidence
Give the application explicit answer, clarification and escalation states. Keep access denials separate so uncertainty handling cannot bypass permissions.
Read articleStart a module with an explicit public contract
Define the operation other features need, then hide the rules and persistence behind it. This gives the module a boundary that can survive internal changes.
Read articleWrite the event intention with the business change
Keep the order update and outbox insert in one transaction. Then give the publisher stable event identities and a clear recovery path for uncertain delivery.
Read articleBind the retry key to the original request
Store the key, account scope and material payload together before executing. Matching retries can reuse the outcome, while conflicting reuse receives an explicit error.
Read articleAdd the new field before requiring it
Introduce a field in stages so older callers remain valid while newer ones adopt it. Define omission and conflict behaviour before changing validation.
Read articleMap each committed effect before defining its compensation
Design the recovery record from the actual side effects. A list of service calls does not tell the coordinator what can safely be corrected after failure.
Read articleCarry tenant identity into queued work
A worker needs an explicit, trusted scope for the job it executes. Save that scope at submission and validate it again according to the workflow's authority policy.
Read articleDefine whether the event is a snapshot or a change
Build an event contract around one interpretation. A consumer needs to know whether it should replace a view, apply a transition or fetch more information.
Read articleMake cache keys describe the result they contain
Include every input that changes a cached response. A resource identifier alone is rarely enough for tenant data, translations or filtered views.
Read articleBackfill without overwriting a newer edit
A migration worker and a live request can touch the same row. Make the update conditional on the state the worker actually read.
Read articleReject an edit based on an old revision
Carry the version a user actually edited into the save request. Compare it atomically with the current record before replacing data.
Read articleStop background writers before the final data sync
A quiet website can still have an active database. Include jobs, callbacks and queued work in the cutover fence.
Read articleProvision one workload environment from a reviewed definition
Make the first environment reproducible before scaling account creation. Include the access, logging and ownership that turn empty infrastructure into a usable workspace.
Read articleRestore into an isolated environment before reconnecting integrations
Build the recovery environment so the application can be tested without repeating production side effects. Reconnect dependencies deliberately after data verification.
Read articlePrepare recovery dependencies before the outage
Reduce the work that must happen under incident pressure. Keep the target's required artefacts, authority and configuration ready through the normal release process.
Read article