| 1 | // Package dbverify is the fleet's restore-drill record: which databases have |
| 2 | // been PROVEN restorable, when, and which could not be decided. |
| 3 | // |
| 4 | // It exists as its own package for the reason internal/usage does — the |
| 5 | // writer (the box-side `carlos ops restore-verify` pass) and any future |
| 6 | // reader (the console) must agree on one shape and one key without either |
| 7 | // importing the other's machinery. |
| 8 | // |
| 9 | // The vocabulary is deliberately three-valued, following |
| 10 | // usage.AppStorage.StoreUnswept: a replica is `ok`, `failed`, or |
| 11 | // `unverified` — never a silent absence and never a 0/false that reads as |
| 12 | // "fine". A pass that could not reach S3 has learned NOTHING about the |
| 13 | // replica, and saying so is the whole point of keeping this record. |
| 14 | package dbverify |
| 15 | |
| 16 | import ( |
| 17 | "context" |
| 18 | "encoding/json" |
| 19 | "errors" |
| 20 | "fmt" |
| 21 | "os" |
| 22 | "time" |
| 23 | |
| 24 | "github.com/carlosframework/platform/internal/release" |
| 25 | ) |
| 26 | |
| 27 | // Key is where the record lives in the deployment bucket, beside |
| 28 | // control/serial and control/registry-snapshot.json. |
| 29 | const Key = "control/restore-verify.json" |
| 30 | |
| 31 | // Verdict is one database's standing. |
| 32 | type Verdict struct { |
| 33 | // OK is true only when the replica was actually restored and passed its |
| 34 | // integrity check on this box, at VerifiedAt. |
| 35 | OK bool `json:"ok"` |
| 36 | // Unverified marks "we could not tell" — an S3 error, a prefix that has |
| 37 | // never been replicated to, or a host the run's cap did not reach. It is |
| 38 | // never a failure and never a pass. |
| 39 | Unverified bool `json:"unverified,omitempty"` |
| 40 | // VerifiedAt is when this verdict was reached — NOT when the replica was |
| 41 | // written. An old timestamp on an ok verdict is itself a finding. |
| 42 | VerifiedAt time.Time `json:"verified_at"` |
| 43 | // Detail carries the failure or the reason it could not be decided. |
| 44 | Detail string `json:"detail,omitempty"` |
| 45 | } |
| 46 | |
| 47 | // Record is the whole box's standing at CheckedAt. |
| 48 | type Record struct { |
| 49 | V int `json:"v"` // 1 |
| 50 | Box string `json:"box"` |
| 51 | CheckedAt time.Time `json:"checked_at"` |
| 52 | Hosts map[string]Verdict `json:"hosts"` |
| 53 | } |
| 54 | |
| 55 | // Merge folds this run's verdicts into the prior record. prior may be nil. |
| 56 | // |
| 57 | // Two rules, both about not overstating coverage: |
| 58 | // |
| 59 | // - A capped run visits only some hosts, so untouched hosts KEEP their |
| 60 | // previous verdict — with its old timestamp, which is what makes staleness |
| 61 | // visible rather than hidden behind a fresh CheckedAt. |
| 62 | // - A host no longer in `fleet` is DROPPED. A retired instance's last "ok" |
| 63 | // would otherwise sit in the record forever, padding the coverage of a |
| 64 | // database that no longer exists. |
| 65 | func Merge(prior *Record, box string, now time.Time, fleet []string, results map[string]Verdict) *Record { |
| 66 | rec := &Record{V: 1, Box: box, CheckedAt: now.UTC(), Hosts: map[string]Verdict{}} |
| 67 | inFleet := map[string]bool{} |
| 68 | for _, h := range fleet { |
| 69 | inFleet[h] = true |
| 70 | } |
| 71 | if prior != nil { |
| 72 | for host, v := range prior.Hosts { |
| 73 | if inFleet[host] { |
| 74 | rec.Hosts[host] = v |
| 75 | } |
| 76 | } |
| 77 | } |
| 78 | for host, v := range results { |
| 79 | if !inFleet[host] { |
| 80 | continue |
| 81 | } |
| 82 | rec.Hosts[host] = v |
| 83 | } |
| 84 | return rec |
| 85 | } |
| 86 | |
| 87 | // Failures counts the verdicts that are an alarm — restored and damaged. |
| 88 | // Unverified rows are deliberately not counted here: they are a coverage |
| 89 | // problem, not a corruption one, and conflating them would either cry wolf or |
| 90 | // train operators to ignore a red run. |
| 91 | func (r *Record) Failures() int { |
| 92 | n := 0 |
| 93 | for _, v := range r.Hosts { |
| 94 | if !v.OK && !v.Unverified { |
| 95 | n++ |
| 96 | } |
| 97 | } |
| 98 | return n |
| 99 | } |
| 100 | |
| 101 | // Unverified counts the verdicts the pass could not decide. |
| 102 | func (r *Record) Unverified() int { |
| 103 | n := 0 |
| 104 | for _, v := range r.Hosts { |
| 105 | if v.Unverified { |
| 106 | n++ |
| 107 | } |
| 108 | } |
| 109 | return n |
| 110 | } |
| 111 | |
| 112 | // Load reads the current record. A missing key is (nil, nil) — a deployment |
| 113 | // that has never run the drill, not an error. |
| 114 | func Load(ctx context.Context, s release.Store) (*Record, error) { |
| 115 | body, _, err := s.Get(ctx, Key) |
| 116 | if errors.Is(err, os.ErrNotExist) { |
| 117 | return nil, nil |
| 118 | } |
| 119 | if err != nil { |
| 120 | return nil, err |
| 121 | } |
| 122 | var rec Record |
| 123 | if err := json.Unmarshal(body, &rec); err != nil { |
| 124 | return nil, fmt.Errorf("parse %s: %w", Key, err) |
| 125 | } |
| 126 | return &rec, nil |
| 127 | } |
| 128 | |
| 129 | // RemoveHost drops host's verdict from the record: a moved instance's |
| 130 | // replica is no longer this fleet's to verify where it was (moves spec |
| 131 | // §6.3, §7.6 local). The load, the delete and the put run inside each CAS |
| 132 | // attempt (moves Plan 2 ruling G31), so a record another writer changed |
| 133 | // between the read and the write is re-read, never overwritten with a stale |
| 134 | // body. No record, or no entry for the host, is nothing to do. |
| 135 | func RemoveHost(ctx context.Context, s release.Store, host string) error { |
| 136 | for attempt := 0; attempt < 3; attempt++ { |
| 137 | body, etag, err := s.Get(ctx, Key) |
| 138 | if errors.Is(err, os.ErrNotExist) { |
| 139 | return nil |
| 140 | } |
| 141 | if err != nil { |
| 142 | return fmt.Errorf("get %s: %w", Key, err) |
| 143 | } |
| 144 | var rec Record |
| 145 | if err := json.Unmarshal(body, &rec); err != nil { |
| 146 | return fmt.Errorf("parse %s: %w", Key, err) |
| 147 | } |
| 148 | if _, ok := rec.Hosts[host]; !ok { |
| 149 | return nil |
| 150 | } |
| 151 | delete(rec.Hosts, host) |
| 152 | next, err := json.Marshal(&rec) |
| 153 | if err != nil { |
| 154 | return err |
| 155 | } |
| 156 | if _, err := s.Put(ctx, Key, next, etag); err == nil { |
| 157 | return nil |
| 158 | } else if !errors.Is(err, release.ErrCASConflict) { |
| 159 | return fmt.Errorf("put %s: %w", Key, err) |
| 160 | } |
| 161 | } |
| 162 | return fmt.Errorf("put %s: lost three CAS races", Key) |
| 163 | } |
| 164 | |
| 165 | // Publish writes rec, replacing whatever is there. Three attempts on a lost |
| 166 | // CAS race, the recipe usage.UpsertEdgeDay and host.putReplacing both use. |
| 167 | func Publish(ctx context.Context, s release.Store, rec *Record) error { |
| 168 | body, err := json.Marshal(rec) |
| 169 | if err != nil { |
| 170 | return err |
| 171 | } |
| 172 | for attempt := 0; attempt < 3; attempt++ { |
| 173 | _, etag, err := s.Get(ctx, Key) |
| 174 | switch { |
| 175 | case errors.Is(err, os.ErrNotExist): |
| 176 | etag = "" |
| 177 | case err != nil: |
| 178 | return fmt.Errorf("get %s: %w", Key, err) |
| 179 | } |
| 180 | if _, err := s.Put(ctx, Key, body, etag); err == nil { |
| 181 | return nil |
| 182 | } else if !errors.Is(err, release.ErrCASConflict) { |
| 183 | return fmt.Errorf("put %s: %w", Key, err) |
| 184 | } |
| 185 | } |
| 186 | return fmt.Errorf("put %s: lost three CAS races", Key) |
| 187 | } |
| 188 | |
amadan