Projects on fyannk

pgToolBox

Mon, 01 Jan 0001 00:00:00 +0000

pgToolBox is a Kubernetes and OpenShift operator. For each PgConsole you declare, it composes a dedicated pod next to one CloudNativePG cluster: an authentication proxy, pgConsole, an embedded pgAdmin, and optionally pgObjectStoreViewer.

One console per cluster, never shared. Dedicated proxy, dedicated pgAdmin, dedicated credentials: isolation wins over resource sharing.
Users are Kubernetes objects. A PgToolBoxUser binds an identity to view, poweruser or dba. A PgToolBoxAccessRequest lets an unknown user ask a dba for access from the proxy’s 403 page.
Any mix of sign-in methods. OIDC, OpenShift OAuth and local accounts at once, so a local account still works when the identity provider is down.
Every exposure type. OpenShift Route, Ingress, Gateway API HTTPRoute or ClusterIP, behind a generated default-deny NetworkPolicy.

It will ship as a Helm chart and an OLM bundle once it leaves pre-alpha.

pgConsole

Mon, 01 Jan 0001 00:00:00 +0000

pgConsole is an operational console for one CloudNativePG Cluster. It shows operator status and conditions, instance pods, recent events, backups, pooler metrics and bounded log tails, and offers a small set of day-2 actions you enable explicitly.

Every claim has a source. Operator-reported, Kubernetes-observed and derived state stay visibly distinct.
Failure stays visible. Broken watches, forbidden reads and stale or truncated data become unknown or stale, never healthy.
A small, guarded mutation surface. Backup, reload, restart and promote are all it can do, each behind access levels, confirmation, CSRF protection, audit and a feature flag.

It reports what CloudNativePG and Kubernetes claim. It does not independently prove replication health, data integrity or restorability.

pgObjectStoreViewer

Mon, 01 Jan 0001 00:00:00 +0000

pgObjectStoreViewer inspects PostgreSQL backup repositories in S3, Azure Blob Storage and GCS. It turns object-store metadata into an inventory, a Barman Cloud backup catalog, WAL continuity diagnostics and conservative recovery coverage.

Read-only by construction. It can list and read bounded metadata. It has no upload, delete or restore path.
WAL-aware. Timelines, duplicates, partial files, and candidate or confirmed gaps.
Honest about gaps in its own view. Incomplete, stale, truncated or unsupported evidence stays unknown.

It reports structural evidence. It does not prove that a restore will succeed. pgBackRest support is planned.

NodeEcho

Mon, 01 Jan 0001 00:00:00 +0000

NodeEcho is a controller that copies selected node state onto the pods running on that node, so a namespace-scoped operator can react to cordons, drains and topology without cluster-wide access to Node objects.

Reflects cordon state and selected taints onto bound pods.
Copies only the node labels and annotations you allow, including custom topology.
Writes nothing when there is nothing to echo, and never overwrites another writer’s keys.

The first use case is CloudNativePG: keeping drain-aware switchovers when the operator cannot read nodes. Upstream CloudNativePG does not consume these labels yet. Not production-ready, and no version is released.

pgAdmin images

Mon, 01 Jan 0001 00:00:00 +0000

Two pgAdmin container images that run without added capabilities, published to ghcr.io/fyannk/pgadmin.

<version> is the official dpage/pgadmin4 release with only the privileged Python capability removed.
<version>-hardened rebuilds that pgAdmin release from source on a fresh base, with refreshed OS and Python dependencies and without Postfix or sudo. latest always points here.

Both listen on 8080 or 8443 by default. pgToolBox runs the hardened image unless told otherwise.

Code Personal

Mon, 01 Jan 0001 00:00:00 +0000

I wanted one clear app for work, with work accounts and work AI assistants, and one for me, with my accounts, my AIs and my own projects. Code Personal is that second app, fully separate from the first: its own name, executable, data folders, URL protocol and icon. It doesn’t conflict with code, visual-studio-code-bin or codium and copies nothing from the work install.

Upstream, barely touched. It builds the stable release of VS Code OSS with only three upstream inputs customized and no TypeScript or CSS patches. No VSCodium patch set is used.
Full-featured. Copilot is built from upstream source and bundled. Extensions come from Open VSX with signature verification kept on, through an adapter for Open VSX’s Ed25519 signatures. A Remote SSH extension is bundled, and it only ever downloads its own server.
Telemetry off by default. VS Code OSS has no Microsoft telemetry endpoint, but bundled Copilot reports usage to GitHub unless telemetry.telemetryLevel is off, so that is the default. You can still opt in.
Built like I’d want a server package built. A pinned Arch container checks upstream every six hours, builds, verifies identity separation and the bundled Copilot, launches the app headless, installs a real Open VSX extension, and insists that a tampered package fails verification. Only then does it publish an Arch package for CachyOS and a .deb, the latter test-installed on Ubuntu 24.04.

The builds are traceable, not claimed to be byte-for-byte reproducible.