Healthie.NET.Dashboard 4.0.0

• net10.0

  • Healthie.NET.Abstractions (>= 4.0.0)

• net8.0

  • Healthie.NET.Abstractions (>= 4.0.0)

Twelve new packages, the schedule model several of them needed, and optimistic concurrency on the
state contract.

**Source-compatible: an application that upgrades without touching its code still compiles and still
behaves as it did.** Nothing public was removed or renamed, and every new interface member is a
defaulted one, so a provider or scheduler written against 3.x keeps working. Two things to know
before upgrading:

- **One binary break.** `HealthieTools`, the MCP read-only tool class, gained a defaulted
 `HealthieMcpOptions?` parameter on its constructor so it could read the page-size option it had
 been ignoring. Calling code still compiles unchanged; an assembly compiled against 3.1.4 and not
 rebuilt does not. Recompiling is enough, and that break is why this is a major rather than a minor.
- **The relational providers alter an existing table on startup** to add the version column
 optimistic concurrency needs. See *Optimistic concurrency* below.

### Added

- **Optimistic concurrency on `IStateProvider`.** A state write can now be made conditional on the
 state not having changed since it was read, closing the lost update that has been documented on
 `CosmosDbStateProvider` since the beginning: a setting changed from the dashboard could be
 overwritten by a check that had read the state first. `GetStateEntryAsync` returns the state with
 the version it was read at, and `TrySetStateAsync` writes only if that version is still current,
 reporting a refused write rather than throwing -- under contention a conflict is expected, not
 exceptional. `UpdateStateAsync` is the read-modify-write retry loop over the pair, and every
 setting on `PulseChecker` goes through it -- as does a check storing its own result, which reads
 and writes the whole state and so used to put every other field back to what it was when the
 check started. Within one process a semaphore hid that; across two replicas sharing one store
 nothing did, and that is the direction the dashboard actually loses to.

 This was expected to need a major release and did not. The new members are defaulted interface
 methods: a provider written against the old two-method interface still compiles, still works, and
 reports `SupportsOptimisticConcurrency == false` rather than pretending. CosmosDB uses ETags,
 the relational providers a version column added to existing tables on startup, and the in-memory
 provider a compare-and-swap.
- **`HealthChanged` on `PulseCheckerStateChangedEventArgs`**, with `PreviousHealth` and
 `CurrentHealth` beside it. `StateChanged` fires on every check -- a stored result always moves the
 execution time -- which was listed as a known issue and is not one: the dashboard redraws from
 exactly those events. What was missing was a way to ask the narrower question, so every handler
 that only cared about a component going down reached into `OldState`/`NewState.LastResult?.Health`
 and worked it out again. The alerting and uptime packages were doing that identically; both now
 read `PreviousHealth` and `CurrentHealth` from the event instead of digging for them.
- **Slack, Microsoft Teams and PagerDuty alert sinks**, in `Healthie.NET.Alerting` beside the
 webhook. Each of those three rejects arbitrary JSON and wants its own shape, so the generic
 webhook never actually reached them without something in between to reshape it -- which its own
 remarks admitted, in the phrase "Teams through a Power Automate flow". They need no dependency the
 package did not already have, so they are sinks rather than a package each. Teams targets the
 Workflows URL and an Adaptive Card, because the Office 365 connectors and the `MessageCard`
 payload they took are retired. PagerDuty resolves the incident it opened rather than raising a
 second one: `Alert.DeduplicationKey` and `Alert.IsRecovery` already existed for exactly that.
- **`Healthie.NET.Redis`.** State is written on every tick of every checker, and Redis is the store
 that does not mind: a relational provider does a round trip to a disk-backed engine for each of
 those writes, and this does one to memory. One hash per checker holding the state, the type it was
 written as, and a version, so the compare and the write are a single Lua script -- Redis runs one
 to completion without interleaving anything else, which is the guarantee a read-then-write cannot
 give. Durability is whatever the Redis is configured for, and the package README says so rather
 than implying more.
- **A startup warning when a surface that can change a checker is reachable without
 authenticating.** `AddHealthieController` still does not require authorization unless asked, and
 the dashboard's `AllowMutations` still defaults to `true` -- both are deliberate and changing
 either would break every application that maps them. What was missing was anyone being told: an
 application that maps one and stops there lets whoever can reach it stop a checker or clear a
 failing streak, which hides an incident rather than reporting one. Logged once, at `Warning`, on
 application start.

 Asked of the endpoints rather than of the flags that built them, so an application that applied
 authorization its own way -- `RequireAuthorization()`, an endpoint group, its own attribute -- is
 not warned at. A warning that fires on correctly secured applications gets filtered out, and then
 it is not there for the one that needs it.
- **Schedules.** `PulseSchedule` says either "every this long" or "on this cron expression", and
 sits alongside `PulseInterval` rather than replacing it. The enum stopped at five minutes, which
 is short of what a certificate-expiry or disk-space check wants. Cron is standard Unix syntax and
 is evaluated in UTC by every scheduler.
- **`Healthie.NET.Postgres`, `Healthie.NET.SqlServer`, `Healthie.NET.Sqlite`** and the
 `Healthie.NET.Relational` engine behind them, which works against any database with an ADO.NET
 driver. PostgreSQL also covers Databricks Lakebase, which is managed PostgreSQL.
- **`Healthie.NET.Hangfire`.** Schedules live in Hangfire's storage, so they survive a restart and
 each occurrence runs on exactly one replica.
- **`Healthie.NET.Coravel`**, for applications already running Coravel. Coravel has no API for
 removing a scheduled job, so Healthie owns the due times and Coravel supplies the tick; the
 package README says so, and says to use the built-in timer if Coravel is not already there.
- **`Healthie.NET.Temporal`.** Schedules live in the Temporal cluster. It needs a cluster and the
 SDK carries a native core, both of which the README weighs against Hangfire and the built-in timer.
- **`Healthie.NET.Checkers`.** HTTP endpoints, TCP ports, TLS certificate expiry, DNS resolution and
 disk space, with no checker code to write. Certificate expiry and disk space report *suspicious*
 before they report unhealthy, because both are gradual and the warning is the useful signal.
- **`Healthie.NET.Alerting`.** Health changes become alerts, delivered to a webhook or your own
 `IAlertSink`, with transition-only firing, recovery notifications and flap suppression. A sink that
 throws, hangs or backs up cannot delay a check, hold its semaphore, or mark anything unhealthy.
- **`Healthie.NET.Uptime`.** Records health transitions rather than checks, so uptime is exact over
 any window and small enough to keep for a year -- the rolling history holds a hundred results,
 which for a one-second checker is a hundred seconds. Time the application was not running is
 reported as unknown rather than counted as either up or down.
- **`Healthie.NET.LeaderElection`.** Runs the checks on one replica at a time. Without it, three
 replicas ask a database three times whether it is healthy and race to write the same state.
- **Metrics and traces.** A `Meter` and an `ActivitySource` named `Healthie.NET`, discovered by
 OpenTelemetry by name with no package to install: check duration, results by health, health
 transitions, and triggers skipped because the previous check had not finished. That last one is
 the only place a checker outlasting its own interval shows up.
- **Bulk state reads.** `IStateProvider.GetStatesAsync` reads every checker in one call, which is
 what a dashboard load and a list request do. Defaulted to the old one-read-per-name behaviour, so
 existing providers are unaffected.
- **State removal.** `IStateProvider.DeleteStateAsync` cleans up after a checker that was renamed or
 removed. Defaulted to refuse rather than to silently do nothing.

- **The dashboard shows what the feature packages know.** Installing one used to change what the
 application did and nothing about the one screen an operator looks at, so uptime, alerts and
 leadership were only visible to whatever the host wired up itself. Each now has a small read-only
 contract in `Healthie.NET.Abstractions` that its package implements, and the board renders the
 panel when the container can resolve it: uptime over the last day and the longest outage in it
 beside the run-based percentage, a drawer of recent alerts saying which reached their sinks and
 which did not, a badge saying whether this replica is the one running the checks, and a button
 that asks the model why a checker has been failing. An application that installs none of them gets
 the board exactly as it was.

 Read-only throughout, so all of it shows under `HealthieUIOptions.AllowMutations = false` -- the
 one exception is asking the model, which is still a read but spends money on the host's account,
 and is gated with the controls that change things.
- **The alert history is persisted and paged.** It is written through the application's own
 `IStateProvider`, so a deployment on CosmosDB, PostgreSQL, SQL Server, SQLite or Redis keeps its
 alerts across a redeploy and one left on the in-memory provider does not. There is no second
 storage contract to configure and no provider had to learn about alerts. `IAlertInsights` reads a
 page at a time and reports the total, because the question asked most often is asked just after a
 restart, about what happened before it -- which is exactly what a last-twenty list throws away.
 Bounded at `HistoryLength`, and the view says how many are kept, of what cap, in which provider.
- **The dashboard shows where alerts are delivered**, with each sink's delivered and failed counts
 and its last error. An application that installed alerting and never configured a sink raises
 alerts that reach nobody, and a healthy-looking list of them was indistinguishable from one that
 was notifying people; sinks are now listed from startup rather than from their first delivery, and
 one that recovers stops being reported as failing.
- **Alerting is configurable from the dashboard**: minimum severity, deduplication window, delivery
 timeout, and whether recoveries alert. Those four and no others, because they are the ones the
 dispatcher reads on every alert -- its queue capacity and history length are fixed when it is
 built, so they are shown as facts rather than offered as controls that would quietly do nothing.
 There is also a test alert, which goes through the real sinks: the only way to find out that a
 webhook URL is wrong is to use it.
- **`AddHealthieMetrics()`** collects the library's own instruments in-process behind
 `IMetricsInsights`, and the dashboard grows a metrics view: checks run, the share that reported
 healthy, transitions, mean and slowest check duration, and results by health. A `MeterListener` on
 the `Healthie.NET` meter, so it reads what is already being emitted and runs alongside an
 OpenTelemetry exporter rather than instead of one. Opt-in, because a listener costs a callback on
 every measurement and an application with an APM has somewhere better to look.

 Overlapped triggers get their own figure. A checker whose check outlasts its own interval looks
 healthy and is quietly running at a fraction of the rate it was asked to, and this is the only
 place that shows.
- **A side menu on the dashboard**, listing an overview, a section per group with its tally and worst
 state, and the alerts, metrics, event-log and about views. Picking a group narrows the list and picking it
 again is the way back; it combines with the search box and the tag filter rather than clearing
 them, and the groups it lists come from the store rather than from the rows currently surviving
 those filters -- a menu that dropped a group because a search had narrowed it away would take away
 the means of getting back to it.

 A rail rather than an overlay drawer, and collapsing to its icons rather than disappearing. This
 package ships no JavaScript of its own, and an off-canvas drawer needs a focus trap to be honest
 about keyboard use. Below 820px it becomes a scrolling strip above the list.
- **The dashboard opens sectioned by group** rather than as one flat list. A group is a partition, so
 the sectioned view answers what is wrong and where at a glance; a flat list of forty checkers asks
 the reader to do that grouping themselves. Checkers with no group collect under one heading, so the
 default hides nothing. The `GROUP` button still switches to the flat list.
- **A cron expression can be set from the dashboard**, beside the interval picker, and through
 `PUT /healthie/{checkerName}/schedule` on the REST API. The scheduler judges the expression before
 anything is stored -- `IPulseScheduler.TryValidateSchedule`, defaulted to accept so an existing
 scheduler is unaffected -- because Cronos, Quartz and Temporal do not agree on cron dialects and
 the only answer worth having is from the implementation that will run it. A refusal carries that
 implementation's own reason and leaves the stored schedule alone; storing first and failing on the
 reschedule would leave a checker that no longer runs and a store that says it should.

 `IPulseChecker.SetScheduleAsync` and `IPulsesScheduler.SetScheduleAsync` are the API behind it. A
 schedule an interval can express exactly is stored as that interval, so only a genuinely custom
 period or a cron expression occupies `PulseCheckerState.Schedule`.

### Fixed

- **The dashboard ignored `PulseSchedule` everywhere it showed a cadence.** It read
 `PulseCheckerState.Interval` for the rate column, for the aggregate checks-per-minute, and for the
 interval picker -- and that field is documented as ignored once `Schedule` is set. So a checker on
 a cron expression advertised a rate it was not running at, was summed into the aggregate at that
 rate, and offered a picker whose every change stored a field nothing reads. The board now reads
 `EffectiveSchedule`, shows a cron expression as one, counts cron checkers separately rather than
 inventing a rate for them, and disables the picker while an expression is in force.
- **`SetIntervalAsync` did nothing to a checker that had a schedule.** The schedule overrides the
 interval, so choosing one left the checker at its old cadence with nothing to say so. It now clears
 the schedule, which is what choosing an interval means.
- **The dashboard's own page did not encode its title.** `MapHealthieUI` builds that one page as a
 string rather than through Razor, which encodes every interpolation for you, so a
 `HealthieUIOptions.DashboardTitle` built from anything the host did not write itself could close
 the title element and open a script one.
- **A dashboard component left its handler behind when it was disposed.** The service is scoped to a
 circuit, so it was assumed the circuit ending released everything; that holds only while the
 dashboard is mounted once per circuit, and a host that routes to it inside its own layout builds a
 new one every time the user navigates back. `IHealthieDashboardService.UnsubscribeFromStateChangesAsync`
 is the missing half, and the component calls it.
- **Two callers scheduling one checker at once could leave a timer nobody could stop.** Installing a
 schedule was "cancel the old one, then start the new one", and only the last one stored was
 reachable; the other kept triggering, could not be unscheduled, and held a linked
 `CancellationTokenSource` that was never disposed.
- **`HealthieMcpOptions.MaxHistoryPageSize` was documented and never read.** `get_check_history`
 clamped to a hard-coded 200 instead, so a host that lowered the option got no such thing.
- **The relational table-name guard admitted a name with a trailing newline.** In .NET `$` matches
 immediately before one, so a table name ending in one passed a check whose own error message says
 it does not. Nothing could be smuggled through it -- a lone trailing newline is whitespace to
 every engine here -- but the guard now ends at `\z` and means what it says.
- **The release workflow granted `contents: write` to every job in it.** Only the one job that
 creates the GitHub release needs it, and the top level now grants nothing but read, so a job added
 later starts with no write access rather than inheriting it.
- **A checker name from a REST route reached the log with its control characters intact.** The
 not-found branch logs a name precisely when it matches nothing, percent-encoded CR and LF arrive
 decoded, and a log sink writing plain text writes them as line breaks.

- A cron schedule whose next occurrence was more than about fifty days out **stopped the checker
 permanently and silently**. `Task.Delay` refuses a longer wait and throws, and that throw is not a
 cancellation, so the scheduler's loop ended with nothing logged. An annual certificate check
 reached it. Long waits are now taken in bounded steps.
- Quartz and the built-in scheduler read the same cron expression in **different timezones** --
 Quartz defaults to the machine's local zone, the built-in scheduler evaluates in UTC. Both are
 pinned to UTC now, matching what the dashboard renders.
- Quartz quietly accepted out-of-range cron fields: `99 99 * * *` produced a trigger firing every
 minute, for ever. Ranges are checked before Quartz sees them.
- Cron expressions were compared as typed, so `0 0 * * MON-FRI` and `0  0  *  *  mon-fri` looked
 like a change to the checker. They are normalized, as tags already were.

### Security

- CodeQL, OpenSSF Scorecard and dependency review run on every pull request, releases carry a
 CycloneDX SBOM, every GitHub Action is pinned to a commit SHA, and every workflow drops to
 deny-by-default permissions.
- Four high-severity advisories resolved: `SQLitePCLRaw.lib.e_sqlite3` (GHSA-2m69-gcr7-jv3q),
 `Newtonsoft.Json` reached through Hangfire (GHSA-5crp-9r3c-p9vr), and `System.Text.Json` in the
 console sample (CVE-2024-30105 and CVE-2024-43485). The solution reports no vulnerable packages.

Full changelog: https://github.com/ivanvyd/Healthie.NET/blob/v4.0.0/CHANGELOG.md