Awesome

rscylla

A reflex stream client for a scylladb CDC log table. It provides an API for consuming scyllaDB CDC logs with at-least-once semantics*. It borrows a lot from scylla-cdc-go.

Usage

// Define your consumer business logic
f := func(ctx context.Context, fate fate.Fate, e *reflex.Event) error {
  fmt.Print("Consuming scylla CDC event", e)
  return fate.Tempt() // Fate injects application errors at runtime, enforcing idempotent logic.
}

// Define some more variables
var keyspace, table, scyllaAddr, consumerName string

// Connect to scylla
cluster := gocql.NewCluster(scyllaAddr)
cluster.PoolConfig.HostSelectionPolicy = gocql.TokenAwareHostPolicy(gocql.DCAwareRoundRobinPolicy("local-dc"))
session, _ := cluster.CreateSession()

// Setup rscylla and reflex
stream := rscylla.NewStream(session, keyspace, table)
consumer := reflex.NewConsumer(consumerName, fn)
cstore := rsql.NewCursorStore() // TODO(corver): Add rscyllla cursor store
spec := reflex.NewSpec(stream.Stream, cstore, consumer)

// Stream forever!
// Progress is stored in the cursor store, so restarts or any error continue where it left off.
for {
  err := reflex.Run(context.Backend(), spec)
  if err != nil { // Note Run always returns non-nil error
    log.Printf("stream error: %v", err)
  }
}

ScyllaDB CDC Overview

ScyllaDB CDC log tables are implemented as normal scylladb tables. A CDC log table contains the append-only log of all changes applied to the base table.

Like all NoSQL DBs of this type, the CDC log data is sharded by the partition key StreamId and sorted by the clustering key Time (and BatchSeqNo).

The CDC log is therefore multiple streams each identified by a StreamID. Ordering is only guaranteed within such a stream.

Note that when the scyllaDB cluster topology changes by a node joining or leaving, a new set of streams are created. This is called "stream generations". Each generation has its own streams and StreamIDs. Generations are sequential in time.

In Kafka terms a CDC log table is a topic and the streams are partitions.

Mapping to Reflex

Reflex was designed for single ordered streams, like mysql event tables or Firehose S3 buckets. A reflex EventID is normally a point in that stream and can therefore be used as a cursor.

Mapping multiple streams to a single reflex stream is therefore a little different.

rsylla basically slices the CDC streams into fixed time windows that are streamed one window at a time. The start of the time window is used as the cursor, not the events themselves. Once all events in the window have been streamed, the cursor is updated to the next window and the cycle begins again.

A rscylla EventID consists of two parts:

• A pointer to the row. The CDC log row's primary key.
• A cursor to the start of the time window. The generation and timestamp.

// eventID is json encoded as the reflex.EventID.
// It points both to the CDC log row itself and the
// time window it was streamed in.
type eventID struct {
  // Cursor of streamed window
  CursorGen  time.Time `json:"cursor_gen"`
  CursorTime time.Time `json:"cursor_time"`

  // CDC log row primary key
  StreamID   string `json:"stream_id"`
  StreamTime string `json:"stream_time"`
  StreamSeq  int `json:"stream_seq"`
}

Limitations

• When starting a stream from a previous EventID, duplicates are to be expected since the cursor points to the start of the time window, not to the specific event in the time window. #at-least-once #idempotent
• The ordering of events in the window is not gauranteed, but the ordering of individual streams is preserved.
• A lag must be specified via the reflex.StreamLag option. Since a time window is queried as a whole, it should only be streamed once it is complete; elapsed with all events present.
• Lagging by less than 30s is considered unsafe and doesn't provide at-least-once semantics. Since scyllaDB is a distributed system, time doesn't have a single source of truth. The timestamp of a CDC log row can be defined by different nodes in the system each with a different clock. The events in a time window is therefore eventually consistent.
• CQL defined timestamps, USING TIMESTAMP 123, may break at-least-once semantics. Client defined timestamps can add events to arbitrary time windows. If the time window has already been streamed, that event is missed.

Features

• Combine an unsafe 1s lag fast stream with a safe +1min sweep stream for logic that requires quick responses to events. The sweeper will pick up any missed events. Most events will be consumed twice though, so ensure idempotency. There is also a risk of concurrent event consumption, since the fast stream can lag.
• Use GetCursor instead of a zero cursor when streaming from the start of a stream. A zero cursor "" is supported but starts at the first generation timestamp of the cluster, the cluster create timestamp. Streaming from this point might result in an initial delay if first events are long after that.
• Use WithConsistency() to tune consistency vs performance. The default is gocql.Quorum.

TODO

• Add support for base row data in the Metadata field depending on type.
• Add support for easier lookup of base table entry from ForeignID.
• Add support for business logic defined Type fields.
• Add reflex.CursorStore implementation backed by a scyllaDB table to rscylla.