Terraform Security and Governance · Topic 3 of 3

Policy Rules and Risky Changes

Terraform Professional (002): layer core and project policy sets, pick enforcement levels, read dangerous plans and apply lifecycle safeguards.

  • 5 min
In a nutshell
  • A policy set uses one framework, and scope can be global, projects, workspaces or tags.
  • Policies evaluate only after a successful plan and before apply.
  • Advisory warns; soft-mandatory can be overridden, as can OPA mandatory when the set allows it; hard-mandatory in policy checks cannot.

What it is

Governance has two layers. Policy sets apply automatic rules to every run, from organization-wide compliance down to rules for one team or project. Plan review and lifecycle settings catch what rules cannot, such as a replacement that would lose data.

Both act before anything changes: policies are checked after the plan and before the apply, and humans read the same plan.

How it works

Layering policy sets

A policy set holds policies from one framework, Sentinel or OPA, and a workspace can have sets from both frameworks applied. The scope you choose decides who is covered.

RequirementScope
Core compliance for everyoneGlobal, with optional exclusions
Rules for one team's projectSelected projects
Rules for specific workspacesSelected workspaces
Rules for a class of workspacesTags

Typical design: a global set holds the rules nobody may skip, such as approved regions and private module use. A project set holds the team's own rules, such as allowed instance sizes.

Keep each team's workspaces in a project so that a project-scoped set also covers future workspaces. Managing sets needs the manage policies permission.

Enforcement levels

FrameworkLevelWhen the policy fails
Sentinel, OPAadvisoryWarning; the run continues
Sentinelsoft-mandatoryRun stops; a permitted user may override
Sentinelhard-mandatoryRun stops; policy checks allow no override
OPAmandatoryRun stops; a permitted user may override if the set allows it

Overriding requires the manage policy overrides permission, and for OPA the policy set must also allow overrides. Audit trails record policy check events, and reading them needs an organization or audit trail token, not a user or team token.

A real example: a Sentinel policy using the tfconfig/v2 import can fail any run whose root module calls do not come from your private registry.

Reading a risky plan

Look past the summary line to the symbols and reasons.

Plan showsMeaningRisk
-/+Destroy, then createDowntime or data loss
+/-Create, then destroyUnique names must not clash
# forces replacementThat argument cannot update in placeOften unexpected
~Update in placeUsually low; still check the attribute

A change of one line can replace a database. Save the plan with terraform plan -out=tfplan, review it, then run terraform apply tfplan, so what you approved is exactly what runs.

Lifecycle and other safeguards

SafeguardWhat it doesLimit
prevent_destroyRejects a plan that destroys the objectRemoving the whole block still destroys it
create_before_destroyBuilds the new object firstBoth objects must be able to coexist
preconditionFails before creating if an assumption is falseChecks only what you wrote
removed with destroy = falseStops managing without destroyingThe object now has no owner
Provider protection, such as deletion_protectionThe API refuses the deleteMust be set before you need it

-target is for exceptional circumstances only: in HCP Terraform, Sentinel sees only the targeted subset and cost estimation is disabled.

Worked example: a database major version upgrade

HashiCorp treats a major RDS upgrade as a risky change. The new engine needs a new parameter group family, and that argument forces replacement of the group, which is still attached to the running instance.

resource "aws_db_parameter_group" "pg16" {
  name_prefix = "app-pg16-"
  family      = "postgres16"

  lifecycle {
    create_before_destroy = true
  }
}

Without create_before_destroy, Terraform would try to replace the group first and fail, because the instance is using it. With it, the order is safe.

Snapshotrecovery pointNew parameter groupInstance upgradedOld group destroyed

Take the snapshot first, and check the plan before approving: the expected summary here is 1 to add, 1 to change and 1 to destroy. Using name_prefix avoids a name clash while both groups exist.

How the exam asks about this

  • "Security requires a rule no workspace can bypass, while each team adds its own." Use a global set with hard-mandatory policies plus project-scoped sets.
  • "A failed cost rule should stop the run, but a manager may approve exceptions." Use soft-mandatory and grant the override permission.
  • "The plan shows the production database with -/+." Stop and find the argument marked forces replacement; do not apply until you understand it.

What to remember for the exam

  • A policy set uses one framework, and scope can be global, projects, workspaces or tags.
  • Policies evaluate only after a successful plan and before apply.
  • Advisory warns; soft-mandatory can be overridden, as can OPA mandatory when the set allows it; hard-mandatory in policy checks cannot.
  • -/+ destroys first; +/- creates first.
  • prevent_destroy does not protect against deleting the resource block from configuration.
  • Apply a saved plan so the reviewed change is exactly what runs.
Watch out

prevent_destroy guards only while the argument remains in the configuration. If someone deletes the resource block, Terraform plans a normal destroy, so pair it with provider deletion protection for critical data.

Common mistakes

Making every rule hard-mandatory. A rule with legitimate exceptions then blocks all work until someone removes it. Keep hard-mandatory for true compliance rules.

Reading only the plan summary. "1 to add, 1 to destroy" may be a harmless swap or your database. Check which resource and which attribute caused it.

Using create_before_destroy with a fixed unique name. The new object cannot be created while the old one holds the name. Use a generated name or prefix.

Check yourself

  • Which scope applies a rule to all current and future workspaces in a team's project?

    A project-scoped policy set, which covers every workspace in the selected project.

  • Who can override a soft-mandatory Sentinel failure?

    A user with the manage policy overrides permission, and the run otherwise stays stopped until it is overridden or discarded.

  • Does prevent_destroy stop a destroy if the resource block is deleted?

    No. The argument must be present in the configuration, so removing the block lets Terraform destroy the object.

Topic 3 of 3 in Terraform Security and Governance

Test yourself on Terraform Security and Governance

The full HashiCorp Terraform Infrastructure Engineer Professional question bank has 600 questions, with timed practice exams and an explanation for every answer.

This study guide is independent and is not affiliated with, endorsed by, or sponsored by HashiCorp. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.