Skip to content

v0.5.0

Choose a tag to compare

@kayladouglas23 kayladouglas23 released this 16 Oct 18:25
· 3 commits to main since this release

Upgrade: hub stacks (DataStack CredsStack WebStack), then NetworkStack and every Connector-<account> stack if you run the private network — see the wiki's Deploy and Upgrade page, "Upgrading to 0.5.0".

  • Bell notices for a change that moves many targets at once are one row, not one per target: when one person gains or loses more than five roles, resources or plugin permissions through a group (a membership or bundle edit) or a cascade, the feed gets a single kind: BUNDLE notice carrying count and, for a group, groupId/groupName ("7,346 access grants · group · FullSystemAccess"). Adding a member to a 7,000-role group used to write 7,000 "access granted" rows for them, the pop-up logic re-raised that backlog as a fresh summary every minute, and POST /notifications/revocations/seen (one update at a time) could not clear it inside the request timeout. Mark-seen now runs 25 updates at a time; the web app folds raised notice keys older than its 500-key cap into a watermark instead of forgetting them. The activity log's per-person rows under such a change collapse the same way: one access.gain/access.lose row per person carrying details.count ("gained 7346 access grants via group FullSystemAccess") instead of one per target; the primary row's affected list is unchanged.
  • Renew sign-in before the 12-hour wall, from anywhere: the user menu's "Session ends" row becomes a "Renew sign-in" button that shows the time left (the end time is its tooltip); the desktop tray's session row gains "Renew sign-in…" (desktop app 0.5.0 or later). All of them start a fresh sign-in while the running session, its credentials and the VPN stay up until the new one lands — nothing is torn down mid-task.
  • Desktop: closing a backend tab no longer loses its address. The connect screen (the + button) lists every deployment the app has connected to at least once under Recently connected, newest first with its project name; click one to reconnect, or its ✕ to forget it after a confirmation. Only that ✕ removes an entry (recents.json in the app's data folder, 20 entries). Reconnecting a saved deployment that was unreachable at startup now keeps its login session instead of starting a fresh one.
  • Large catalogs (#39): the portal now holds a 4,000-account, 16,000-role catalog without the pages or the API growing with it. The catalog (accounts, roles, group names) lives in an in-memory index per API container, rebuilt only when a catalog write moves a version row, and every catalog-sized list — my access, the admin Accounts page, the request picker, the users directory, a group's bundle — serves one paged contract (?q=&filter[key]=a,b&sort=&limit=&cursor= → {items,nextCursor,total,facets}; without limit each route answers its old whole-list shape, so existing clients keep working). Access is still decided in one place before any page is cut; the index lists, it never grants.
  • Dashboard (#45): a user who holds an account an admin kept off the dashboard gets a sticky search over everything they hold — results drop down as role rows with the card's actions (star, Console, Get credentials, the desktop auto-login bolt) — above Signed in now (roles with running credentials), Main, Favorites and Recently vended, each heading pinned while its cards scroll, and a table of the Other accounts (the roles outside Main and Favorites, five to start and more as you scroll; GET /accounts gains the main and favorite facets and a sparse main: false on each account for it); everyone else keeps the plain page unchanged.
  • Lists in panels and dialogs (a user's or group's roles, a service user's roles, a group editor's picked roles) show five rows, with Show more / Show less and a search box as soon as any are folded away, instead of thirty and "Show more"; the grant-roles dialog is wider and its dimension filter no longer opens off the dialog's edge. New account flag Show in the Main section (on by default; main: false on POST/PATCH /admin/accounts), GET /accounts/home, PUT /preferences/favorites/{account} (50 favorites, synced across devices), recents from credential and console vends. The admin Accounts page has a deployment-wide Dashboard layout choice (PUT /admin/dashboard-layout): automatic (the rule above), sections for everyone, or the plain list for everyone.
  • Account dimensions (#47), opt-in: accountDimensions in the project config stores tenant / project / stage on each account when it is registered, read off its name (<tenant>-<project>-<stage>; a pipeline may pass dims instead), admin-editable on the account form and never re-derived by a rename. They filter the dashboard search, the admin Accounts page, every account picker (request form, role form, grant dialogs) and the group editor, which gains add one role across many accounts (POST /admin/groups/{id}/roles/bulk, with a dry-run count). The role card's silhouette follows the stage through the config's stage map (stages); without the block nothing changes. A deployment turning it on backfills existing rows with scripts/backfill-dimensions.mjs (deployment README "Account dimensions"). The dashboard's search and browse results can be grouped by one dimension (a Group by choice beside the filter), and registering a role offers the groups that already bundle that role in the account's sibling accounts (same tenant and stage, GET /admin/groups/suggest).
  • Groups at scale (#43, #44): POST /admin/groups/{id}/roles and DELETE …/roles/{account}/{role} add or remove one role, written only if absent, so a pipeline registering accounts never races another writer (the full-list PATCH is read-modify-write); the add answers with the change ({groupId, name, role, added}), not the bundle, and neither call reads the whole bundle — a 4,000-role group answers in well under a second; the access diff behind a bundle edit reads only the touched targets; the group list caps each bundle at 50 roles with roleCount and GET /admin/groups/{id}/roles pages the rest; one yearly review per group (its clock = the latest approval or bundle change) replaces one per bundled role; the users directory caps each user's role and resource lists at 50 with totals, GET /admin/users/{id} carries everything.
  • Registration is idempotent (#43): POST /admin/accounts and POST /admin/accounts/{a}/roles answer 200 with the stored row when the same registration is sent again, 409 only when a field differs — a pipeline can retry any step. The admin forms still refuse a duplicate.
  • Target-account template (#43): TargetAccountStack / targetAccountTemplate() synthesize one account-agnostic CloudFormation template (native AWS::IAM::OIDCProvider, sub pin through AWS::AccountId) for StackSets or Control Tower, so a 4,000-account deployment does not need one CDK stack per account.
  • Pickers (#46): every place that listed the whole catalog in a <select> or a checkbox wall (request form, add-role form, group form, the grant dialogs on user, group and service-user panels) is a typeahead over one page at a time.
  • Access report (#46): POST /admin/access-report/jobs builds the CSV in the background (a second Lambda on the API bundle, 15-minute budget, private reports bucket with a 7-day expiry) and GET …/jobs/{id} hands back a 15-minute download link; the Users page's export uses it. The inline GET /admin/access-report remains.
  • Reads bounded by the catalog (#41, #44): BatchGetItem retries unprocessed keys and runs in parallel chunks; the admin bell reads the bare directory; the setup checklist counts from the index; the reconciler scans only what it needs.
  • Security Hub S3.9 and CloudFront.5: every hub bucket (web app, OIDC issuer, plugin assets) sends S3 server-access logs, and both CloudFront distributions send standard logs, into one new AccessLogsBucket in DataStack (90-day expiry, kept on teardown when prod is on). That bucket is the one that logs nowhere; AWS's own S3.9 text says to suppress its finding. Rides a deploy of all three stacks.
  • Private network: the ALB writes its access logs to a bucket of its own in NetworkStack (Security Hub ELB.5): SSE-S3, TLS-only, public access blocked, bucket-owner-enforced, objects expire after 90 days, its own access logs go to the hub's shared access-log bucket under s3/alb-logs/; the bucket is kept on a prod teardown and emptied and deleted otherwise. Rides a NetworkStack deploy.
  • Private network: ECS.20 passes on every network task definition. The control-server task is three non-root containers with no shared task volume: the init container renders the config and the self-signed TLS pair next to the database on the EFS mount (the access point performs every file operation as uid 0, whatever user the container runs as), the server container serves from there with its Unix socket on a per-container tmpfs and runs the two CLI bootstrap steps itself once the socket answers, so the bootstrap-user and bootstrap-key sidecars are gone and Custom::HeadscaleApiKey reads the key from the headscale container's own log stream (headscale-<site>/headscale/<task>); the publish-cert step reads the certificate from EFS. The connector task's two containers are non-root too, each with its own tmpfs, and the minted auth key crosses from the aws-cli container to the Tailscale container over the task's localhost instead of a shared file. The control server moves to the upstream -debug image tag (same Headscale, plus busybox for the start script). Why the layout changed: a Fargate task volume is root-owned 0755 when the task starts whatever mode the image gives the path, so 0.4.0's non-root containers could only share a volume a root init container had chowned, which kept ECS.20 failing on that container and on the connectors. Rides a NetworkStack deploy (each control server restarts once; the EFS directory gains config.yaml, tls.key and tls.crt) and a deploy of every connector stack (each connector re-registers once). Needs Fargate's tmpfs support (platform version LATEST, available in all Regions since January 2026) and aws-cdk-lib 2.263 or later.
  • Bell: temporary plugin permissions now warn before they end and once more after, like roles and network resources. New routes GET /plugins/expiry and POST /plugins/expiry/seen; the notice names the plugin and the account.
  • Removing an account now also denies its open plugin reviews and deletes a group that bundled nothing but that account's plugin permissions, as the role cascade already did.
  • PUT /admin/plugins/{pluginId}/accounts/{accountNumber}/grants answers 404 for a user that is neither in the pool nor a registered service user; DELETE still takes any id, so a deleted user's row stays revocable. Every plugin answer carries Cache-Control: no-store, 204s and errors included.
  • The activity log names the plugin on a revoke (it only did on a grant).
  • Plugin assets are uploaded with Cache-Control: max-age=31536000, immutable (their prefix carries the content hash), matching the app bundle.
  • Catalog provider: a construct rename no longer tombstones the record the new resource just wrote (the record remembers the logical id that owns it; CloudFormation's cleanup Delete for the old id is ignored). Provider tests now cover a retried Create, rollback after a failed install and after a failed update, and a reinstall over expired rows.
  • CDK tests pin that no method under a plugin path is unauthenticated, that the catalog record depends on the API stage, that no Function URL exists, and the feature's on → empty → off lifecycle.

Full changelog: CHANGELOG.md · npm: @fntio/infra-client