Terms that actually come up in projects, briefly explained.
Every entry ends with the catch the textbook leaves out.
Legacy and modernization
A pattern for replacing a legacy system step by step: new functionality grows around it, the old system shrinks, until nothing of it is left.
The deferred effort from earlier shortcuts in code or architecture. It does not disappear, it simply falls due later and at a higher price.
A translation layer between new and old, so that the quirks of the legacy system do not seep into new code.
A switch in the code that turns functionality on and off at runtime, without shipping a new release.
An application that ships as one unit but has clearly separated modules with real boundaries inside.
A procedure for database changes without downtime: expand first, switch second, remove the old thing last.
The number of people who would have to disappear before a system can no longer be developed or operated. For many legacy systems it is one.
The point from which a version no longer receives security updates. From then on every known hole stays open permanently.
Greenfield means building on an empty site. Brownfield means building while an existing system keeps running around you.
Switching a system on a fixed date, usually over a weekend, with no intermediate state and only one attempt.
Code without tests that earns money. The definition by Michael Feathers is useful because it names the real problem: missing safety, not age.
Tests that record what a system does today, rather than checking what it should do. The safety net before the first change.
The share of code executed during a test run. A number that says a lot about gaps and little about quality.
Improving the internal structure of code without changing its externally visible behaviour.
An application built, deployed and operated as a single unit.
An application is split into several independently deployable services, each with its own storage and lifecycle.
Moving existing data into a new system, usually the underestimated and most schedule-critical part of a replacement.
The interface is designed and agreed first, implementation and user interface are built against it afterwards.
AWS and cost
The prepared foundation of an AWS environment: accounts, networking, permissions, logging and cost allocation, before the first application moves in.
The most detailed cost report AWS offers: every resource, every hour, every pricing component as raw data instead of a chart.
The practice of treating cloud cost as an ongoing engineering decision rather than a monthly invoice that lands in accounting.
Two ways to pay considerably less than on-demand rates in exchange for a one or three year usage commitment.
Bringing resources down to the size actually needed, rather than the one chosen when they were created.
The service through which resources in private subnets reach the internet. Billed per hour and per gigabyte processed.
Charges for data leaving an AWS region, an availability zone or the cloud. Inbound data is usually free.
A binding scheme that assigns every resource to its team, its environment and its cost centre.
A question catalogue from AWS for assessing an architecture along six pillars, from operational excellence to sustainability.
Spare capacity that AWS sells at up to 90 percent off and can reclaim at short notice.
Capacity follows load automatically, so that neither users wait nor idle resources get paid for.
Storage tiers by access frequency, from instantly available to archive, at very different prices.
A system moves to the cloud almost unchanged, without adapting to cloud characteristics.
A system is deliberately adapted to the platform during the move, without being rewritten.
An operating model where the provider handles provisioning and scaling and only actual usage is billed.
The service that runs individual functions without your own servers, triggered by events or calls.
Your own virtual network in AWS, where subnets, routes and access rules are defined.
Access management in AWS: who or what may perform which action on which resource.
Managed relational databases on AWS: backup, patching and failover are handled by the provider.
The AWS delivery network: content is cached at locations close to the user.
Backend and operations
Two numbers for the bad day: how long recovery may take (RTO), and how much data may be lost in the process (RPO).
Two complete environments side by side: the new version runs quietly, traffic is switched over, and on failure you switch back.
One system serves several customers on the same infrastructure, without the data of one appearing for another.
A system runs in several geographic regions, so the failure of one region does not end the service, and users have shorter distances.
A operation is idempotent when running the same request several times has the same result as running it once.
A protective mechanism that stops calling a failed service for a while, instead of running into every timeout.
An upper bound on requests per period and caller, so that individual users cannot take a system down for everyone else.
The property of a system to reveal from the outside what is happening inside, including for questions nobody asked in advance.
A measurement (SLI), a self-imposed target for it (SLO), and a contractual promise with consequences (SLA).
Services exchange events instead of calling each other directly. The sender does not know who is listening.
Infrastructure is described as versioned code instead of being clicked together in a web console.
A system loses individual features under partial failure instead of going down entirely.
A method for calls between services with a declared interface and binary transport, considerably leaner than JSON over HTTP.
A way to run operations across several services when there is no shared transaction: every step has a compensating step.
A new version goes to a small share of traffic first. If the numbers hold, the share increases.
Data is split across several databases by a key, when a single one no longer carries the volume or the load.
An application together with its dependencies in a self-contained package that runs the same everywhere.
A system that distributes containers across machines, restarts them, scales them and manages their reachability.
An automated path from change to deployment: build, test, ship, with no manual steps in between.
A distributor that spreads incoming requests across instances and takes failed ones out of rotation.
Results are stored temporarily so that expensive computations or queries do not have to repeat.
A buffer between sender and receiver: messages wait until someone processes them.
A service in front that accepts requests, validates them and forwards them to the services behind.
A style for HTTP interfaces: resources with clear addresses, standard methods and stateless calls.
A reversed call: instead of polling, you get notified by the other system as soon as something happens.
An additional structure that lets the database find rows without reading the whole table.
Controlled disruption in a running system, to find out whether the safeguards actually hold.
A short instruction for a recurring operational case: what to check, what to do, whom to call.
AI in production
A language model is handed relevant excerpts from your own data before answering, instead of relying on its training.
An open standard through which language models reach tools and data sources: one shared interface instead of one integration per vendor.
An attack that hides instructions inside data a language model processes, so it does something other than intended.
Number sequences that represent the meaning of a text. Similar content sits close together in vector space, even without shared words.
Technical boundaries around a language model that define what may go in, what may come out, and what it can trigger.
A repeatable procedure for measuring whether answers from a language model are getting better or worse.
The amount of text a language model can consider at once for a request, measured in tokens.
An answer that sounds plausible and is factually wrong. The model does not invent deliberately, it strings likely words together.
An existing model is further trained on your own examples, to hit style, format or domain language reliably.
A language model that does not only answer but calls tools over several steps, judges intermediate results and continues.
The running cost per request to a language model, billed by tokens in and tokens out.
A flow in which a person confirms, corrects or rejects at defined points before an action takes effect.
A database that stores vectors and finds similar entries quickly, the basis for search by meaning.
Splitting documents into sections that can be searched individually and handed to a model.
A second step in retrieval: the first hits are reordered by a more accurate model before they reach the language model.
Deliberately phrasing instructions to a language model so that results are reliable rather than accidentally good.
The base instruction that precedes every user input and sets role, boundaries and output format.
Instead of text, a model returns a structured request for which tool to call with which parameters.
Models whose weights are freely available and which can therefore be run yourself, in your data centre or your own cloud.
The European regulation for artificial intelligence, classifying applications by risk and attaching obligations to that class.
Process and automation
A recurring workflow is taken over by software instead of being carried between systems by people.
The point where data changes format or system and a human has to step in.
A transaction runs end to end without human intervention, from intake to booking.
A service that drives multi-step workflows, holds state, handles retries and continues in a defined way after failures.
Data is extracted from source systems, transformed and loaded into a target, usually on a schedule.
Connecting a system to inventory management, so that stock, prices, orders and documents are not maintained twice.
Software operates existing applications the way a human would, through their user interface rather than an API.
Security and operational risk
Every identity gets exactly the permissions its task requires, and none beyond that.
Managing passwords, keys and credentials outside of code and configuration files.
No access is trusted merely because it comes from the internal network. Every request is verified.
An immutable record of who performed which action and when.
Data is encrypted both while being transferred and while being stored.
The organized handling of security updates: detect, assess, apply, evidence.
A commissioned attack on your own system, to find weaknesses before somebody else does.
Diese Website nutzt Google Analytics, um den Besuch anonym auszuwerten. Ihre Daten werden nur nach ausdrücklicher Einwilligung erhoben.Datenschutzerklärung