Terraform Pro 002 flashcards

Scalable Terraform configuration design

• Read requirements for timing, failure mode and persistence — those words decide which Terraform mechanism fits
• No default means required; nullable = false rejects null on required inputs and makes null fall back to the default on optional ones
• optional(type, default) fills omitted and null object attributes; a variable default is replaced wholesale by any caller value and never merges
• Variable defaults must be literals: derive conventions in locals (coalesce for overridable values) — modules cannot read their own outputs
• Environment differences are explicit inputs set by each root module, never workspace-name or environment-name checks inside the module
• Key for_each by stable, unique identifiers known at plan time; apply-time IDs are values, not keys; fixed timestamps belong in state (time_static)
• Validation (1.9+ may reference other variables), preconditions and postconditions error; check blocks and their scoped data sources only warn; postconditions never roll back
• sensitive redacts display and propagates, but values remain in state and saved plans; outputs built from sensitive values must be marked sensitive
• Ephemeral values and write-only arguments are never persisted; root outputs cannot be ephemeral; bump the _wo_version argument to send a new write-only value
• Commit .terraform.lock.hcl (lock every platform with terraform providers lock -platform), use >= required_version in shared modules, and pair prevent_destroy with provider deletion protection

• Last loaded wins: TF_VAR_ < terraform.tfvars < *.auto.tfvars (by filename) < -var/-var-file in order; the default is only a fallback.
• Undeclared names: TF_VAR_ is silently ignored, a .tfvars file only warns, -var is an error — and every one of these sources sets root variables only.
• Object conversion silently drops undeclared attributes; optional(type, default) fills attributes that are omitted or null.
• A nullable variable passed null gets null; nullable = false makes null take the default, or fail when there is no default.
• Locals are module-scoped across files, referenced as local.name, and never visible to child or parent modules; values go down through arguments and up through outputs.
• terraform output shows root outputs only; terraform console evaluates any expression, and -scope evaluates inside a child module instance.
• Sensitivity follows the value through functions, locals and module outputs: mark root outputs sensitive, and unmark only what becomes a for_each key.
• for_each takes a map or a set of strings with keys known at plan; instances are addressed by key, form a map (no bare splat), and renaming a key is destroy and create.

• merge() lets the last argument win and is shallow: pass values that must not be overridden last, and merge nested maps explicitly.
• lookup() defaults a missing key, try()/can() catch a missing attribute path, coalesce() skips null and empty strings.
• Map keys and sets of strings iterate in lexical order; sort() converts numbers to strings and sorts them as text.
• element() wraps around, list[i] does not, and slice() excludes its end index and always returns a list.
• A for expression that produces duplicate map keys fails unless you add ... to group values into lists.
• replace() is literal unless the pattern is wrapped in /slashes/; cidrsubnets() aligns each range to its own size.
• optional(type, default) substitutes the default for omitted or null attributes, applies nested defaults top-down and drops undeclared attributes.
• Dynamic blocks use the block label (or iterator) not each, cannot generate lifecycle or provisioner, and use cond ? [] : [x] for an optional block.
• Preconditions and postconditions stop the run (only postconditions can use self); check block assertions only warn.
• templatefile() sees only the variables you pass it; provider functions use provider::<local name>::<fn> and need Terraform 1.8+ plus a provider version that ships the function.

• A resource or data block without a provider argument uses the unaliased default configuration; aliases are selected with provider = <local name>.<alias>.
• If every block for a provider is aliased, Terraform creates an implied empty default configuration that unaliased resources and child modules use.
• Child modules inherit default configurations only; aliased ones are never inherited, and setting providers cancels default inheritance, so list everything the child needs.
• Modules that need alternate configurations declare configuration_aliases; callers map them, one level at a time, and one configuration can fill several names.
• Modules with their own provider blocks cannot use count, for_each or depends_on, and any provider configuration must stay until its resources are destroyed.
• Every module declares its own required_providers; an undeclared local name falls back to the hashicorp namespace.
• Provider arguments must be known during plan: split configurations when they depend on computed attributes, and read the other side with a data source.
• The Terraform MCP server's registry toolset needs no token; registry-private and terraform need a valid TFE_TOKEN (plus TFE_ADDRESS for Terraform Enterprise).
• Omitting --toolsets loads all (token-gated toolsets register only with a valid TFE_TOKEN); the default keyword selects the registry toolset alone.
• ENABLE_TF_OPERATIONS (default false) gates destructive tools; the token's permissions bound everything; the hosted remote server reaches US HCP Terraform organizations only and is never destructive.

Terraform state and backend operations

• Backend blocks accept literals only — no variables, locals, data sources or terraform.workspace — and a configuration has one backend, never alongside a cloud block.
• Vary backend values per environment with -backend-config files or KEY=VALUE pairs at init; later options override earlier ones and the block. Only init accepts -backend-config.
• Backend configuration values, credentials included, persist in .terraform/terraform.tfstate and in saved plans, so supply credentials through environment variables or roles.
• After a backend change, -migrate-state copies state and -reconfigure starts empty; -force-copy answers the prompts (and implies -migrate-state), and -input=false without it aborts.
• S3 locking is use_lockfile (off by default): a .tflock object at the state key that needs Get, Put and Delete permissions. Keep dynamodb_table alongside it until every job runs the new config.
• Versioning recovers state, a customer managed KMS key needs Encrypt, Decrypt and GenerateDataKey, and only an access control such as a bucket-policy deny keeps secrets in state from other readers.
• Non-default S3 workspaces live at <workspace_key_prefix>/<workspace>/<key> (prefix env:), and so do their lock files.
• The cloud block maps to one workspace by name or many by tags (never both); TF_CLOUD_* and TF_WORKSPACE fill only omitted attributes, and an empty cloud {} is still required.
• terraform_remote_state reads root outputs only, defaults to the default workspace and needs full state access; HCP sharing is set on the producer, while tfe_outputs is governed by its token.

• state list prints addresses (filter by prefix or -id), state show prints one instance's attributes, show takes a file path not an address, and output reads root outputs only.
• for_each keys are quoted strings attached to the segment that repeats; wrap the whole address in single quotes on Unix shells.
• show -json, output -json and output -raw print sensitive values in plain text; sensitive_values in the JSON marks what to strip.
• terraform console evaluates expressions against state and holds a state lock while open; exit it instead of force-unlocking.
• state pull is the backup; state push refuses a different lineage or a higher remote serial, and -force overwrites without merging.
• apply -refresh-only accepts drift into state only; update the configuration too, or the next normal apply reverts it or recreates a deleted object.
• A normal plan and apply reverts drift; -refresh=false and state-only views cannot see it.
• removed blocks (destroy = false) forget objects in a reviewed plan but cannot target one instance key; state rm needs the block deleted as well.
• moved blocks go from old to new address; count to for_each needs one per instance. Prefer -replace over taint, and untaint a healthy object tainted by a failed provisioner.
• Health assessments need remote or agent execution and a successful latest run; org enforcement overrides workspaces; fix drift by plan and apply or by changing configuration.

• Only installing a different binary changes the Terraform release; required_version checks it, the lock file tracks providers only, and init -upgrade moves providers and modules.
• The running CLI must satisfy every required_version in the module tree; an unmet one, even in a nested or registry module, stops Terraform before it acts.
• ~> lets only the right-most stated component grow: ~> 1.12.0 means patches only, ~> 1.12 allows any later 1.x. Pre-releases match only exact constraints.
• Reusable modules declare a minimum (>=) at the latest feature they use; root modules bound both ends.
• terraform providers shows which module declares each provider constraint, terraform modules -json lists every module call, terraform graph shows ordering, and state pull's terraform_version shows what last wrote state.
• terraform graph only knows references and depends_on; add depends_on for run-time-only dependencies, but prefer references over module-level depends_on.
• Upgrade terraform_remote_state readers no later than the producer, and publish renamed outputs alongside the old name until consumers switch.
• A saved plan applies only with the same Terraform release and the same lock selections, so pin one exact release for plan and apply and use plain init in both.
• Roll dev to stage to prod, core and providers in separate changes; plan never persists state, and a direct jump between v1.x releases is supported.
• Downgrades are not guaranteed: snapshot state before the first apply and roll back by restoring it (state push -force) with the old version.

Infrastructure lifecycle and operations

• CLI workspaces share one configuration, backend and credentials; they fit temporary parallel copies, not environments that need separate credentials, access controls or state storage.
• workspace select fails on a missing workspace unless -or-create is set; workspace new fails on an existing one and can seed state with -state.
• workspace delete refuses the current workspace and one still tracking resources; -force leaves real objects unmanaged, and default can never be deleted.
• terraform.workspace is default when nothing else is selected; indexing a map with an unknown workspace fails, so use lookup() with a fallback.
• TF_WORKSPACE is for non-interactive use, and while it is set, workspace select refuses to switch.
• A root per environment gives each its own backend, credentials, constraints, lock file and module pins, so changes are promoted by bumping each root in turn.
• Switch a shared directory between environments with init -reconfigure, never -migrate-state; give concurrent jobs separate TF_DATA_DIR values.
• Scale by splitting roots along rate of change and ownership; -target is for exceptional use and hides drift.
• Lower -parallelism (default 10) for API rate limits; set it per command with TF_CLI_ARGS_plan/apply, which also replaces TFE_PARALLELISM on agents.
• Stacks: components in .tfcomponent.hcl, one deployment per environment in .tfdeploy.hcl; ungrouped deployments get <name>_default with no auto-approval; retire with destroy = true.

• terraform apply FILE executes exactly the saved plan: no approval prompt, recorded variable values (sensitive ones included), no re-planning and no planning options such as -var, -target or -refresh=false.
• Run options such as -parallelism, -lock-timeout and -input=false are accepted with a saved plan because they change how it runs, not what it contains.
• Any write to state after planning makes a saved plan stale; the only fix is a new plan and a new review, never refresh, state rollback or waiting on the lock.
• Hand plans between jobs with the whole initialised working directory (.terraform and the lock file), the same Terraform and provider versions and the same OS/architecture; never init -upgrade before applying.
• Plan files and their JSON rendering hold sensitive values in cleartext; restrict and expire them like secrets.
• With -detailed-exitcode, 0 = no changes, 1 = error, 2 = changes; scripts must branch on 1 and 2 separately.
• Policy gates test whether change.actions contains "delete" (covering both replace encodings), check after_unknown before reading after, and never rely on action_reason.
• -target includes dependencies, never dependents, is for exceptional cases only, and should be followed by a full plan.
• Apply is not transactional: completed changes stay in state, failed creates are tainted, the first Ctrl-C stops gracefully, and recovery is a fresh reviewed plan.
• In HCP Terraform, plain CLI plans and pull-request plans are speculative; saved cloud plans (CLI 1.6+) and CLI applies work only in workspaces without a linked VCS repository.

• Import blocks change state only when their reviewed plan is applied; terraform import writes state immediately with no plan, which can leave shared state ahead of the merged code.
• The to address must match a declared managed resource instance exactly: module.<call>.<type>.<name>, [n] for count, ["key"] for for_each, and the key on whichever block has for_each. Data sources cannot be imported.
• A for_each import block must reproduce the resource's own instance keys; iterating the same map as the resource is the simplest way.
• The import id must be known at plan time: variables, locals and data sources work; attributes of resources created in the same run do not. id and identity are mutually exclusive.
• A safe import plan is import-only. An extra update or replacement means the resource block differs from reality: fix the block, not the lifecycle settings.
• terraform plan -generate-config-out writes resource blocks with literal values to a new file only; the import block's provider argument is allowed only when generating configuration.
• terraform query with list blocks in .tfquery.hcl discovers unknown objects; limit defaults to 100 and include_resource adds attributes; -generate-config-out writes import and resource blocks.
• create_before_destroy, prevent_destroy and ignore_changes take literal values only; for per-environment settings use two count-toggled resource blocks. replace_triggered_by takes managed-resource references, and conditions take expressions.
• create_before_destroy needs unique-name handling, spreads to the resources it depends on and stops destroy provisioners; replace_triggered_by takes managed resources or attributes, with terraform_data for plain values and each.key for per-instance scope.
• action_trigger runs provider actions on create and update events, drift fixes included; -invoke runs one action alone; moved blocks refactor module calls, and a shared module ships its own moves.

• Classify the error first: configuration, initialization, state, platform API or provider plugin each have a different fix.
• TF_LOG turns logging on and sets the level; TF_LOG_PATH only redirects it to a file and does nothing alone. TF_LOG=JSON is TRACE in JSON and its format is not stable.
• TF_LOG_PROVIDER isolates plugin and API traffic; TF_LOG_CORE isolates Terraform's own graph and evaluation work.
• For HCP Terraform remote runs, set TF_LOG as a workspace environment variable, leave TF_LOG_PATH unset, and remember Sensitive hides the value, not the trace output.
• terraform init -backend=false then terraform validate checks code offline; it never calls provider APIs, so platform errors appear only at plan or apply. Use -json for CI annotations.
• A locked version outside a new constraint, or a bad release you excluded with !=, needs terraform init -upgrade; commit the lock file and never hand-edit it.
• Lock files built through a mirror hold hashes for one platform; run terraform providers lock -platform=... for every OS and architecture and commit.
• depends_on never breaks a cycle; restructure references, and use terraform graph -type=plan -draw-cycles to see large ones.
• Stale lock: confirm nothing runs, force-unlock, then plan. Failed state save: terraform state push errored.tfstate.
• Remove resources before their provider configuration; apply force_destroy before destroy; untaint a healthy object a timed-out create left tainted; import an 'already exists' object only when no other state owns it.

• terraform output -raw handles only strings, numbers and booleans; use -json (with jq) for lists, maps and objects
• terraform output -json without a name returns {name: {sensitive, type, value}}; with a name it returns the bare value
• sensitive = true hides values in plan, apply and the unnamed output listing only; state, named output, -json and -raw expose them in plain text
• terraform_remote_state exposes only root outputs but requires read access to the whole state snapshot, and copies what it reads into the consumer's state
• Child module values reach other projects only through a root output, and only after an apply records it in state
• Publishing values to SSM, Consul, a ConfigMap or DNS grants access per value; direct data source lookups remove state coupling but must match exactly one object
• In HCP Terraform use tfe_outputs (values is sensitive wholesale; nonsensitive_values is the safe subset) and run triggers, which fire on a successful apply and share no data
• The external data source needs a JSON object of strings, reports errors via stderr and a non-zero exit, and runs on every plan, so it must have no side effects
• Terraform combines every module's provider constraints and needs one version satisfying all; reusable modules set minimums, root modules use ~>, pre-releases need an exact version
• The lock file pins providers only: commit it, upgrade with init -upgrade where the change is made, add platforms with terraform providers lock -platform, and pin modules exactly (Git by commit SHA)

CI/CD and automation workflows

• terraform fmt -check -recursive is the read-only formatting gate; add -diff for useful logs. Without -recursive, subdirectories are skipped; without -check, fmt rewrites files.
• The plugin cache is not concurrency safe and is used only for packages that match lock-file checksums; offline runners need -plugin-dir or a filesystem_mirror-only provider_installation.
• setup-terraform's default wrapper exposes stdout, stderr and exitcode outputs but corrupts piped output; set terraform_wrapper: false for redirection.
• TF_TOKEN_<host> encodes periods as _ and hyphens as __, and beats credentials written by terraform login.
• VCS triggers: always (any file), specified paths (globs or prefixes, prefixes need a working directory, include local modules), or Git tags (every matching tag, regardless of branch).
• API-driven configuration versions: archive the directory's contents; auto-queue-runs defaults to true; speculative versions can never be applied.
• Run tasks block only when mandatory, and timeouts count as failures; inspect changes at post-plan or pre-apply.
• Auto-apply covers UI, API and VCS runs, not CLI runs (-auto-approve); auto-approve only non-critical environments and keep one outstanding plan per state.
• Blue/green and canary: weights move traffic instantly at the load balancer, toggles move capacity; create, verify, shift, then decommission. DNS weights wait on TTLs, forward-action stickiness delays returning users.
• Promote exact, tested artifacts (exact module versions, pinned image IDs or HCP Packer channels) and roll back by applying the last good configuration; HCP state rollback changes state only.

• AWS dynamic credentials need TFC_AWS_PROVIDER_AUTH=true and a role ARN, set as environment variables; no STS call at all means the switch is missing, a denied call means the trust policy is wrong
• The sub claim is organization:ORG:project:PROJECT:workspace:WORKSPACE:run_phase:PHASE; wildcard it only with StringLike and always check aud and the organization
• Renaming or moving an organization, project or workspace changes sub (the ID claims do not); a custom AWS audience must be added to both the OIDC provider and the trust policy
• PLAN and APPLY role variables beat the RUN role, which is only a fallback; each role must trust its own run phase
• Azure federated credentials match exactly (one per workspace per phase); Google Cloud uses assertion.sub conditions and the numeric project number
• Tagged configurations need their own PROVIDER_AUTH_<TAG>=true, live under aliases["<TAG>"], and force you to map every provider, the default one included; TFC_DEFAULT_ values only fill gaps
• Workspace variables override variable sets; more specific sets beat global sets; priority sets beat everything, including -var
• Vault-backed credentials need Vault dynamic credentials first, then TFC_VAULT_BACKED_ variables, with secrets-engine reads removed from code
• On the command line, the last -var or -var-file wins; auto files load from the root directory only, in name order, above terraform.tfvars(.json) and TF_VAR_
• Maps and objects are never merged across variable sources; combine separate variables with merge() instead

Terraform security and governance

• sensitive = true redacts output; it never keeps a value out of state — only ephemeral sources feeding write-only arguments do
• CLI-driven runs upload the directory filtered by .terraformignore (not .gitignore) and load uploaded *.auto.tfvars files
• Sensitive workspace variables are write-only for everyone: overwrite to rotate, delete and recreate to change the key or flag
• Ephemeral values are allowed only in write-only arguments, provider blocks, locals, ephemeral blocks, ephemeral variables/outputs and provisioners; anywhere else the plan fails
• Write-only values are never diffed: bump the _wo_version argument to push a rotated secret
• manage_master_user_password lets RDS own and rotate the password in Secrets Manager; it conflicts with password and password_wo, and master_user_secret_kms_key_id encrypts the secret
• With Vault dynamic provider credentials, drop token/address from the provider, set skip_child_token = true, use renewable short-TTL tokens, and use ephemeral or Vault-backed credentials rather than secret-generating data sources
• Trust policies must check aud and a sub that includes the organization; use StringLike for wildcards and pin each of the plan and apply roles to its own run_phase
• GitHub OIDC needs id-token: write; an environment in the job changes sub to environment:NAME
• Every fixed workspace permission set reads full state — use custom Read outputs only and tfe_outputs to share outputs safely; team tokens for runs, the organization token for setup

• Organization baselines go in a globally scoped policy set; team rules go in extra sets scoped to a project, workspace or tag.
• Every policy set in scope is evaluated independently, so a team's own set can never relax a global rule; exemptions are exclusions on the set that fails.
• Exclude the unit that matches the exempt group: a workspace, a project, or a tag that cuts across projects.
• Sentinel levels are advisory, soft-mandatory and hard-mandatory; OPA has advisory (the default) and mandatory.
• Manage policy overrides lets a team continue overridable failures run by run; Manage policies is for authoring and does not grant overrides.
• A hard-mandatory policy-check failure has no override: fix the code or the policy and queue a new run.
• Policy evaluations (Agent-type Sentinel, all OPA) allow runtime pinning and a set-level override option but cannot read cost estimates.
• Policies run after a successful plan and before apply; an errored plan never reaches them and never applies.
• Test change.actions contains "delete" to catch deletes and both replacement orders; tfrun.is_destroy only sees destroy plans.
• prevent_destroy disappears with its resource block; pair a mandatory delete policy with provider deletion protection, and split workspaces to shrink the blast radius.

Advanced module usage and registry practices

• Write a module when resources are used together and repeated, or when standards must be enforced; avoid thin wrappers around one resource type.
• Reuse across repositories with independent upgrades needs versioned releases that callers pin, not a local path or a branch reference.
• Scope modules by encapsulation, privileges and volatility; prefer small composable modules over one module full of create_* flags.
• Pass dependencies in as inputs; use data-only modules to hide how shared values are retrieved.
• moved blocks go in the module that can see both addresses, carry instance keys where for_each lives, and let a shared module refactor itself for every caller.
• The contract is inputs, outputs, provider requirements and resource addresses; locals and internal wiring are private, and narrow outputs keep the contract small.
• Required inputs have no default; optional() defaults replace both omitted and null attributes; specific types catch caller mistakes on entry.
• Modules declare required_providers with a source and a >= minimum and contain no provider blocks.
• Breaking changes (new required inputs, type changes, removed outputs or moved blocks) need a major version; Terraform 1.15 deprecated warns without failing.

• Public registry addresses are NAMESPACE/NAME/PROVIDER; private ones add the hostname: HOSTNAME/ORG/NAME/PROVIDER; // selects a nested module
• version works only with registry sources, and any change to source or version needs terraform init; the lock file covers providers, not modules
• On Terraform 1.15, module source and version can use const = true variables and locals built only from them
• Private registry credentials come from CLI credentials blocks or TF_TOKEN_ variables (periods as _, hyphens as __), using user or team tokens — never organization tokens
• Registry docs are per version: match the docs to the provider or module version you actually run
• Public registry: public GitHub repo named terraform-<PROVIDER>-<NAME>, standard structure, x.y.z tags with optional v; private monorepos need a Source Directory and a tag prefix per module
• Deleting a Git tag does not remove a registry version — delete it in the registry
• A run defaults to command = apply and creates real infrastructure, destroyed in reverse run order when the file ends; plan runs can only assert on known values
• expect_failures covers custom conditions only, in the run’s own command; test-file variables beat -var and tfvars, run-level beats file-level
• Mocks generate computed values during apply unless override_during = plan; parallel runs need parallel = true, no cross-references and separate state (state_key)

HCP Terraform and Terraform Enterprise operations

• Agent execution = private network with the full HCP Terraform run pipeline; agents connect outbound only.
• Local execution keeps state storage, locking and sharing but drops variables, variable sets, notifications, health assessments, policies and run tasks.
• Scoping an agent pool doesn't detach existing users; rotate agent tokens make-before-break; any agent may run any phase.
• CLI runs ignore workspace auto-apply (use -auto-approve), cannot apply in VCS-linked workspaces, and Ctrl-C after apply starts stops only the logs.
• Every fixed role reads full state; only a custom role with Read outputs only hides secrets. Project Maintain administers workspaces; project Admin administers the project.
• A lock blocks applies and uploads by others; deleting a workspace never destroys its infrastructure.
• Generic webhooks: set a token, verify X-TFE-Notification-Signature (HMAC SHA-512); enabling needs a 2xx test response.
• Continuous validation needs Terraform 1.3+ and checks or conditions; assessments never change infrastructure and yield to runs.
• HCP Terraform bills the hourly peak of managed resources (workspaces + Stacks, minus null_resource and terraform_data); partial hours count in full.
• Cost estimation = cloud spend, off by default, skipped for -target; tfrun.cost_estimate prior/proposed are before/after monthly totals (decimal strings), delta = proposed minus prior.

• cloud block tags map a directory to every workspace carrying all listed tags and allow CLI workspace commands; name pins one workspace and disables them.
• TF_CLOUD_ORGANIZATION, TF_CLOUD_HOSTNAME and TF_CLOUD_PROJECT fill omitted cloud block settings; TF_WORKSPACE selects but never creates a workspace; TFE_* variables belong to the tfe provider.
• New workspaces share state with an empty specific list (pre-April-2021 ones default to organization-wide); tfe_outputs reads outputs without full state access, and its nonsensitive_values can drive for_each.
• Run triggers live on the downstream workspace, fire after a successful source apply, and auto-apply only when Auto-apply run triggers is on.
• Local execution mode keeps state storage and locking but drops workspace variables, variable sets, policies, run records and the CLI version check.
• A workspace version constraint resolves to the newest matching release; use exact versions where changes must be reviewed, and manage terraform_version per environment in tfe_workspace.
• Upgrade order: widen required_version, plan-only run on the new version, change the workspace version, then an empty apply to rewrite state; terraform-version is valid only on plan-only runs.
• sentinel test reads test/<policy>/*.hcl; a case without a test block expects a pass; run mocks need only a successful plan, expire after an hour via the API, and are not guaranteed scrubbed.
• OPA tests need input shaped as {plan, run}; the JSON plan endpoint returns the raw plan, needs workspace admin with a user or team token, and redirects to a short-lived URL.
• VCS-connected policy sets update on push (policies_path, policy_update_patterns) and refuse API uploads; unconnected sets take tar.gz versions or tfe_slug; pin runtimes with policy_tool_version (OPA rejects latest; Sentinel Agent mode has no cost data).

Yes — the whole Terraform Pro 002 deck is free and needs no account, as is every Terraform Pro 002 revision lesson. Create a free account to save progress and drill practice questions; full-length mock exams are a one-time unlock per certification.