| 1 | package console |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "strings" |
| 7 | "testing" |
| 8 | "time" |
| 9 | |
| 10 | "github.com/carlosframework/platform/internal/metrics" |
| 11 | "github.com/carlosframework/platform/internal/release" |
| 12 | ) |
| 13 | |
| 14 | // sentMail is one captured sendApp call. The token is captured so a test can |
| 15 | // assert each operator is mailed with THEIR OWN grant, never a shared one. |
| 16 | type sentMail struct { |
| 17 | token, to, subject, body string |
| 18 | } |
| 19 | |
| 20 | // alertFixture is a console fixture with a captured sendApp. err makes |
| 21 | // sends fail; failFor narrows the failure to one recipient, which is how |
| 22 | // the per-recipient retry is pinned. |
| 23 | type alertFixture struct { |
| 24 | *consoleFixture |
| 25 | sent []sentMail |
| 26 | err error |
| 27 | failFor string |
| 28 | } |
| 29 | |
| 30 | // newCapturedFixture wires the capture and nothing else — no operators, so |
| 31 | // a test can build its own recipient shape. |
| 32 | func newCapturedFixture(t *testing.T) *alertFixture { |
| 33 | t.Helper() |
| 34 | fx := newConsoleFixture(t) |
| 35 | af := &alertFixture{consoleFixture: fx} |
| 36 | fx.srv.sendApp = func(_ context.Context, token, to, subject, body string) (string, error) { |
| 37 | af.sent = append(af.sent, sentMail{token, to, subject, body}) |
| 38 | if af.err != nil && (af.failFor == "" || af.failFor == to) { |
| 39 | return "", af.err |
| 40 | } |
| 41 | return "msg-1", nil |
| 42 | } |
| 43 | return af |
| 44 | } |
| 45 | |
| 46 | // newAlertFixture is the common shape: two operators, only one of whom ever |
| 47 | // granted the console send access. |
| 48 | func newAlertFixture(t *testing.T) *alertFixture { |
| 49 | t.Helper() |
| 50 | af := newCapturedFixture(t) |
| 51 | for _, addr := range []string{"ops-a@keymail.dev", "ops-b@keymail.dev"} { |
| 52 | if err := af.st.SetOperator(addr, true); err != nil { |
| 53 | t.Fatal(err) |
| 54 | } |
| 55 | } |
| 56 | if err := af.st.SetGrantToken("ops-a@keymail.dev", "grant-a", time.Now()); err != nil { |
| 57 | t.Fatal(err) |
| 58 | } |
| 59 | return af |
| 60 | } |
| 61 | |
| 62 | func (af *alertFixture) writeAlert(t *testing.T, st metrics.AlertState) { |
| 63 | t.Helper() |
| 64 | b, err := json.Marshal(st) |
| 65 | if err != nil { |
| 66 | t.Fatal(err) |
| 67 | } |
| 68 | // Publishing is a CAS write (create-only on ""), so re-publishing a box's |
| 69 | // state — which is what a fire→clear transition is — needs the etag. |
| 70 | key := metrics.AlertKey(st.Box) |
| 71 | etag := "" |
| 72 | if _, cur, err := af.rel.Get(context.Background(), key); err == nil { |
| 73 | etag = cur |
| 74 | } |
| 75 | if _, err := af.rel.Put(context.Background(), key, b, etag); err != nil { |
| 76 | t.Fatal(err) |
| 77 | } |
| 78 | } |
| 79 | |
| 80 | func (af *alertFixture) pass(t *testing.T) { |
| 81 | t.Helper() |
| 82 | af.sent = nil |
| 83 | if err := af.srv.alertPass(context.Background()); err != nil { |
| 84 | t.Fatal(err) |
| 85 | } |
| 86 | } |
| 87 | |
| 88 | func firingDisk(firing bool) metrics.AlertState { |
| 89 | rs := metrics.RuleState{Firing: firing, Value: 0.95} |
| 90 | if firing { |
| 91 | rs.FiredAt = time.Date(2026, 8, 2, 10, 0, 0, 0, time.UTC) |
| 92 | rs.Since = rs.FiredAt |
| 93 | } else { |
| 94 | rs.Value = 0.80 |
| 95 | } |
| 96 | return metrics.AlertState{ |
| 97 | V: 1, |
| 98 | Box: "box-a", |
| 99 | UpdatedAt: time.Date(2026, 8, 2, 10, 5, 0, 0, time.UTC), |
| 100 | Rules: map[string]metrics.RuleState{"disk": rs}, |
| 101 | } |
| 102 | } |
| 103 | |
| 104 | // TestAlertPassFiresOnceThenGoesQuiet is the spec's core promise: a keymail |
| 105 | // on the TRANSITION, to the operators who hold a grant, and silence while |
| 106 | // the alert stays in the state the console already reported. |
| 107 | func TestAlertPassFiresOnceThenGoesQuiet(t *testing.T) { |
| 108 | af := newAlertFixture(t) |
| 109 | af.writeAlert(t, firingDisk(true)) |
| 110 | |
| 111 | af.pass(t) |
| 112 | if len(af.sent) != 1 { |
| 113 | t.Fatalf("first pass sent %d keymails, want 1: %+v", len(af.sent), af.sent) |
| 114 | } |
| 115 | got := af.sent[0] |
| 116 | if got.to != "ops-a@keymail.dev" || got.token != "grant-a" { |
| 117 | t.Fatalf("sent to %q with token %q, want ops-a@keymail.dev/grant-a", got.to, got.token) |
| 118 | } |
| 119 | if got.subject != "[carlos] box-a: disk alert" { |
| 120 | t.Fatalf("subject = %q", got.subject) |
| 121 | } |
| 122 | // "95%", not "0.95": the mail carries the rule's own unit, the same as |
| 123 | // the operator dashboard's alerts table (see alertValueText). |
| 124 | for _, want := range []string{"disk", "box-a", "95%", "2026-08-02T10:00:00Z"} { |
| 125 | if !strings.Contains(got.body, want) { |
| 126 | t.Fatalf("body is missing %q:\n%s", want, got.body) |
| 127 | } |
| 128 | } |
| 129 | if strings.Contains(got.body, "grant-a") { |
| 130 | t.Fatal("body leaked the grant token") |
| 131 | } |
| 132 | |
| 133 | // Recorded: the same state on the next pass says nothing. |
| 134 | firing, known, err := af.st.AlertSeen("box-a", "disk", "ops-a@keymail.dev") |
| 135 | if err != nil || !known || !firing { |
| 136 | t.Fatalf("AlertSeen = %v,%v,%v — want true,true,nil", firing, known, err) |
| 137 | } |
| 138 | af.pass(t) |
| 139 | if len(af.sent) != 0 { |
| 140 | t.Fatalf("second pass sent %d keymails, want 0: %+v", len(af.sent), af.sent) |
| 141 | } |
| 142 | } |
| 143 | |
| 144 | // TestAlertPassClears covers the other half of the transition: the alert |
| 145 | // going away is news too. |
| 146 | func TestAlertPassClears(t *testing.T) { |
| 147 | af := newAlertFixture(t) |
| 148 | af.writeAlert(t, firingDisk(true)) |
| 149 | af.pass(t) |
| 150 | |
| 151 | af.writeAlert(t, firingDisk(false)) |
| 152 | af.pass(t) |
| 153 | if len(af.sent) != 1 { |
| 154 | t.Fatalf("clear pass sent %d keymails, want 1: %+v", len(af.sent), af.sent) |
| 155 | } |
| 156 | if af.sent[0].subject != "[carlos] box-a: disk cleared" { |
| 157 | t.Fatalf("subject = %q", af.sent[0].subject) |
| 158 | } |
| 159 | af.pass(t) |
| 160 | if len(af.sent) != 0 { |
| 161 | t.Fatalf("pass after clear sent %d keymails, want 0", len(af.sent)) |
| 162 | } |
| 163 | } |
| 164 | |
| 165 | // TestAlertPassRetriesAfterSendFailure: a failed send leaves the transition |
| 166 | // unrecorded, so the next tick delivers it. A missed page is worse than a |
| 167 | // duplicate one. |
| 168 | func TestAlertPassRetriesAfterSendFailure(t *testing.T) { |
| 169 | af := newAlertFixture(t) |
| 170 | af.writeAlert(t, firingDisk(true)) |
| 171 | af.err = context.DeadlineExceeded |
| 172 | |
| 173 | af.pass(t) |
| 174 | if len(af.sent) != 1 { |
| 175 | t.Fatalf("failing pass attempted %d sends, want 1", len(af.sent)) |
| 176 | } |
| 177 | if _, known, err := af.st.AlertSeen("box-a", "disk", "ops-a@keymail.dev"); err != nil || known { |
| 178 | t.Fatalf("AlertSeen recorded despite a send failure (known=%v, err=%v)", known, err) |
| 179 | } |
| 180 | |
| 181 | af.err = nil |
| 182 | af.pass(t) |
| 183 | if len(af.sent) != 1 { |
| 184 | t.Fatalf("retry pass sent %d keymails, want 1", len(af.sent)) |
| 185 | } |
| 186 | if firing, known, err := af.st.AlertSeen("box-a", "disk", "ops-a@keymail.dev"); err != nil || !known || !firing { |
| 187 | t.Fatalf("AlertSeen = %v,%v,%v after a successful retry", firing, known, err) |
| 188 | } |
| 189 | } |
| 190 | |
| 191 | // TestAlertPassRetriesOnlyTheFailedOperator is why alert_seen is keyed per |
| 192 | // recipient: one operator's dead grant must not re-page every healthy |
| 193 | // operator on the fleet, tick after tick. |
| 194 | func TestAlertPassRetriesOnlyTheFailedOperator(t *testing.T) { |
| 195 | af := newAlertFixture(t) |
| 196 | if err := af.st.SetGrantToken("ops-b@keymail.dev", "grant-b", time.Now()); err != nil { |
| 197 | t.Fatal(err) |
| 198 | } |
| 199 | af.writeAlert(t, firingDisk(true)) |
| 200 | af.err = context.DeadlineExceeded |
| 201 | af.failFor = "ops-b@keymail.dev" |
| 202 | |
| 203 | af.pass(t) |
| 204 | if len(af.sent) != 2 { |
| 205 | t.Fatalf("first pass attempted %d sends, want 2: %+v", len(af.sent), af.sent) |
| 206 | } |
| 207 | if firing, known, err := af.st.AlertSeen("box-a", "disk", "ops-a@keymail.dev"); err != nil || !known || !firing { |
| 208 | t.Fatalf("ops-a AlertSeen = %v %v %v — a healthy delivery was not recorded", firing, known, err) |
| 209 | } |
| 210 | if _, known, err := af.st.AlertSeen("box-a", "disk", "ops-b@keymail.dev"); err != nil || known { |
| 211 | t.Fatalf("ops-b AlertSeen recorded despite a failed send (known=%v, err=%v)", known, err) |
| 212 | } |
| 213 | |
| 214 | af.err = nil |
| 215 | af.pass(t) |
| 216 | if len(af.sent) != 1 || af.sent[0].to != "ops-b@keymail.dev" { |
| 217 | t.Fatalf("retry pass sent %+v, want exactly one mail to ops-b", af.sent) |
| 218 | } |
| 219 | if af.sent[0].token != "grant-b" { |
| 220 | t.Fatalf("retry used token %q, want ops-b's own grant", af.sent[0].token) |
| 221 | } |
| 222 | af.pass(t) |
| 223 | if len(af.sent) != 0 { |
| 224 | t.Fatalf("pass after the retry sent %d keymails, want 0", len(af.sent)) |
| 225 | } |
| 226 | } |
| 227 | |
| 228 | // TestAlertPassNoMailableOperator pins the one path where an alert is |
| 229 | // genuinely dropped: nobody holds a grant, so there is nothing to send and |
| 230 | // the transition is recorded against the synthetic recipient so it does not |
| 231 | // replay every minute forever. The honest cost — documented, and asserted |
| 232 | // here rather than left to be discovered — is that an operator who gains a |
| 233 | // grant afterwards is never told about that transition. |
| 234 | func TestAlertPassNoMailableOperator(t *testing.T) { |
| 235 | af := newCapturedFixture(t) |
| 236 | if err := af.st.SetOperator("ops-b@keymail.dev", true); err != nil { |
| 237 | t.Fatal(err) |
| 238 | } |
| 239 | af.writeAlert(t, firingDisk(true)) |
| 240 | |
| 241 | af.pass(t) |
| 242 | if len(af.sent) != 0 { |
| 243 | t.Fatalf("sent %d keymails with no grant anywhere: %+v", len(af.sent), af.sent) |
| 244 | } |
| 245 | if firing, known, err := af.st.AlertSeen("box-a", "disk", ""); err != nil || !known || !firing { |
| 246 | t.Fatalf("synthetic AlertSeen = %v %v %v, want true true nil — the drop was not recorded", firing, known, err) |
| 247 | } |
| 248 | |
| 249 | // The grant arrives after the fact. The alert is still firing, but the |
| 250 | // console has already recorded the transition as handled: no catch-up. |
| 251 | if err := af.st.SetGrantToken("ops-b@keymail.dev", "grant-b", time.Now()); err != nil { |
| 252 | t.Fatal(err) |
| 253 | } |
| 254 | af.pass(t) |
| 255 | if len(af.sent) != 0 { |
| 256 | t.Fatalf("a newly-granted operator was retroactively paged: %+v", af.sent) |
| 257 | } |
| 258 | // But the NEXT transition reaches them normally. |
| 259 | af.writeAlert(t, firingDisk(false)) |
| 260 | af.pass(t) |
| 261 | if len(af.sent) != 1 || af.sent[0].to != "ops-b@keymail.dev" { |
| 262 | t.Fatalf("clear after the grant arrived sent %+v, want one mail to ops-b", af.sent) |
| 263 | } |
| 264 | } |
| 265 | |
| 266 | // TestAlertPassQuietFirstSighting: a rule the console has never seen that is |
| 267 | // NOT firing is recorded silently. Nobody wants a keymail saying the disk is |
| 268 | // fine. |
| 269 | func TestAlertPassQuietFirstSighting(t *testing.T) { |
| 270 | af := newAlertFixture(t) |
| 271 | af.writeAlert(t, firingDisk(false)) |
| 272 | |
| 273 | af.pass(t) |
| 274 | if len(af.sent) != 0 { |
| 275 | t.Fatalf("quiet first sighting sent %d keymails, want 0: %+v", len(af.sent), af.sent) |
| 276 | } |
| 277 | if firing, known, err := af.st.AlertSeen("box-a", "disk", "ops-a@keymail.dev"); err != nil || !known || firing { |
| 278 | t.Fatalf("AlertSeen = %v,%v,%v — want false,true,nil", firing, known, err) |
| 279 | } |
| 280 | } |
| 281 | |
| 282 | // TestAlertPassSurvivesJunk: an unparsable object under alerts/ is skipped, |
| 283 | // and the boxes either side of it are still processed. A watcher that dies |
| 284 | // on one bad file is a watcher that stops paging. |
| 285 | func TestAlertPassSurvivesJunk(t *testing.T) { |
| 286 | af := newAlertFixture(t) |
| 287 | if _, err := af.rel.Put(context.Background(), "alerts/boxes/box-junk.json", []byte("{not json"), ""); err != nil { |
| 288 | t.Fatal(err) |
| 289 | } |
| 290 | af.writeAlert(t, firingDisk(true)) |
| 291 | |
| 292 | af.pass(t) |
| 293 | if len(af.sent) != 1 { |
| 294 | t.Fatalf("pass sent %d keymails, want 1: %+v", len(af.sent), af.sent) |
| 295 | } |
| 296 | } |
| 297 | |
| 298 | // TestAlertPassEmptyBucket: no alerts/ prefix at all is a no-op, not an |
| 299 | // error — a fleet that has never published an alert is the normal case. |
| 300 | func TestAlertPassEmptyBucket(t *testing.T) { |
| 301 | fx := newConsoleFixture(t) |
| 302 | fx.srv.sendApp = func(context.Context, string, string, string, string) (string, error) { |
| 303 | t.Fatal("sent a keymail with no alerts published") |
| 304 | return "", nil |
| 305 | } |
| 306 | if err := fx.srv.alertPass(context.Background()); err != nil { |
| 307 | t.Fatal(err) |
| 308 | } |
| 309 | } |
| 310 | |
| 311 | // TestNewServerDefaultsSendApp: production wiring goes through the console's |
| 312 | // own keymail client; nothing else in NewServer starts a goroutine. |
| 313 | func TestNewServerDefaultsSendApp(t *testing.T) { |
| 314 | fx := newConsoleFixture(t) |
| 315 | if fx.srv.sendApp == nil { |
| 316 | t.Fatal("NewServer left sendApp nil — the watcher would never deliver") |
| 317 | } |
| 318 | var _ release.Store = fx.rel |
| 319 | } |
| 320 | |
| 321 | func geoipAlert(code float64, firing bool) metrics.AlertState { |
| 322 | rs := metrics.RuleState{Firing: firing, Value: code, FiredAt: time.Date(2026, 10, 6, 10, 0, 0, 0, time.UTC)} |
| 323 | return metrics.AlertState{V: 1, Box: "box-a", UpdatedAt: time.Date(2026, 10, 6, 10, 5, 0, 0, time.UTC), |
| 324 | Rules: map[string]metrics.RuleState{"geoip": rs}} |
| 325 | } |
| 326 | |
| 327 | // Code review R4-3: a geoip reason that changes while firing is mailed. |
| 328 | func TestAlertPassMailsAGeoIPReasonChange(t *testing.T) { |
| 329 | af := newAlertFixture(t) |
| 330 | step := func(st metrics.AlertState, wantMails int, wantBody string) { |
| 331 | t.Helper() |
| 332 | af.writeAlert(t, st) |
| 333 | af.pass(t) // resets af.sent |
| 334 | got := af.sent |
| 335 | if len(got) != wantMails { |
| 336 | t.Fatalf("mails = %d, want %d: %+v", len(got), wantMails, got) |
| 337 | } |
| 338 | if wantBody != "" && !strings.Contains(got[0].body, wantBody) { |
| 339 | t.Fatalf("body lacks %q: %s", wantBody, got[0].body) |
| 340 | } |
| 341 | } |
| 342 | step(geoipAlert(2, true), 1, "expires within 9 days") |
| 343 | step(geoipAlert(2, true), 0, "") |
| 344 | step(geoipAlert(3, true), 1, "expired and removed") |
| 345 | step(geoipAlert(0, false), 1, "") |
| 346 | // A later episode, and a return to an earlier reason within one: each |
| 347 | // is news (code review R5 on R4-3). |
| 348 | step(geoipAlert(2, true), 1, "expires within 9 days") |
| 349 | step(geoipAlert(3, true), 1, "expired and removed") |
| 350 | step(geoipAlert(2, true), 1, "expires within 9 days") |
| 351 | step(geoipAlert(2, true), 0, "") |
| 352 | } |
| 353 | |