package models import ( "fmt" "github.com/emvi/logbuch" "github.com/mitchellh/hashstructure/v2" "strings" "time" ) type Heartbeat struct { ID uint64 `gorm:"primary_key" hash:"ignore"` User *User `json:"-" gorm:"not null; constraint:OnUpdate:CASCADE,OnDelete:CASCADE;" hash:"ignore"` UserID string `json:"-" gorm:"not null; index:idx_time_user"` Entity string `json:"entity" gorm:"not null; index:idx_entity"` Type string `json:"type"` Category string `json:"category"` Project string `json:"project"` Branch string `json:"branch"` Language string `json:"language" gorm:"index:idx_language"` IsWrite bool `json:"is_write"` Editor string `json:"editor" hash:"ignore"` // ignored because editor might be parsed differently by wakatime OperatingSystem string `json:"operating_system" hash:"ignore"` // ignored because os might be parsed differently by wakatime Machine string `json:"machine" hash:"ignore"` // ignored because wakatime api doesn't return machines currently UserAgent string `json:"user_agent" hash:"ignore"` Time CustomTime `json:"time" gorm:"type:timestamp; index:idx_time,idx_time_user" swaggertype:"primitive,number"` Hash string `json:"-" gorm:"type:varchar(17); uniqueIndex"` Origin string `json:"-" hash:"ignore"` OriginId string `json:"-" hash:"ignore"` CreatedAt CustomTime `json:"created_at" gorm:"type:timestamp" swaggertype:"primitive,number" hash:"ignore"` // https://gorm.io/docs/conventions.html#CreatedAt } func (h *Heartbeat) Valid() bool { return h.User != nil && h.UserID != "" && h.User.ID == h.UserID && h.Time != CustomTime(time.Time{}) } func (h *Heartbeat) Augment(languageMappings map[string]string) { maxPrec := -1 // precision / mapping complexity -> more concrete ones shall take precedence for ending, value := range languageMappings { if ok, prec := strings.HasSuffix(h.Entity, "."+ending), strings.Count(ending, "."); ok && prec > maxPrec { h.Language = value maxPrec = prec } } } func (h *Heartbeat) GetKey(t uint8) (key string) { switch t { case SummaryProject: key = h.Project case SummaryEditor: key = h.Editor case SummaryLanguage: key = h.Language case SummaryOS: key = h.OperatingSystem case SummaryMachine: key = h.Machine case SummaryBranch: key = h.Branch } if key == "" { key = UnknownSummaryKey } return key } func (h *Heartbeat) String() string { return fmt.Sprintf( "Heartbeat {user=%s, entity=%s, type=%s, category=%s, project=%s, branch=%s, language=%s, iswrite=%v, editor=%s, os=%s, machine=%s, time=%d}", h.UserID, h.Entity, h.Type, h.Category, h.Project, h.Branch, h.Language, h.IsWrite, h.Editor, h.OperatingSystem, h.Machine, (time.Time(h.Time)).UnixNano(), ) } // Hash is used to prevent duplicate heartbeats // Using a UNIQUE INDEX over all relevant columns would be more straightforward, // whereas manually computing this kind of hash is quite cumbersome. However, // such a unique index would, according to https://stackoverflow.com/q/65980064/3112139, // essentially double the space required for heartbeats, so we decided to go this way. func (h *Heartbeat) Hashed() *Heartbeat { hash, err := hashstructure.Hash(h, hashstructure.FormatV2, nil) if err != nil { logbuch.Error("CRITICAL ERROR: failed to hash struct – %v", err) } h.Hash = fmt.Sprintf("%x", hash) // "uint64 values with high bit set are not supported" return h }