RT @CockroachDB: You shouldn’t have to guess how your database will behave under chaos. With Performance under Adversity, you will get proo….

I haven't contributed to GIMP in decades (literally), yet it still holds a special place in my heart. It is really amazing to see the community continue to develop it.

😎.

8/8 Analogies help us understand complexity—and I'd love to hear yours. Does this resonate with you? What analogies help you make sense of our rapidly evolving relationship with AI?.

7/8 If you're still on the fence or just unaware, I encourage you: get in the car. The journey is better together, and there's so much to discover.

6/8 We're actively learning how to collaborate, constantly reshaping each other as we navigate this new terrain. The future feels thrilling, unknown, and filled with potential, but our choices now matter immensely.

5/8 And here's the fascinating part: the AI "car" itself is continually transforming beneath us. Yesterday it was a Model T, today a sleek Chevy, tomorrow perhaps a Tesla with autonomous driving. What's next—a flying car? A starship?.

4/8 That car is AI. Excited yet cautious, we all climb in. We learn by doing—steering, braking, accelerating—aware of risks but inspired by the incredible possibilities unfolding.

3/8 Now, imagine a group of engineers who've carefully designed and built a car—powerful, intriguing, yet complex enough that even they aren't sure exactly how it will behave. Together with broader society, they begin exploring how to responsibly guide and control this vehicle.

2/8 Humanity—and each of us individually—has always augmented our abilities through tools. We've moved from simple sticks and wheels to sophisticated machines, each transforming how we interact with the world.

1/8 I'm not an AI luminary—I'm a technologist who loves exploring complex ideas through analogies. Lately, I've been reflecting on our collective journey with AI. Here's one analogy that resonates deeply with me.

Slack falls silent.Seek an unbreakable core.@CockroachDB awaits the call.

@CockroachDB If you enjoyed this thread, take a look at which provides more details on the Go maps Swiss Tables implementation.

8/8 Columnar blocks are a big leap from rowblk. Inspired by columnar systems (common in OLAP), we adapted it for OLTP. Huge thanks to the @CockroachDB Pebble team—future columnar improvements are on the horizon!.

7/8 Due to its use of MVCC, CockroachDB frequently has to skip over different versions of a key using a NextPrefix operation. The colblk.Bitmap ( records when the prefix changes allowing next prefix determination without looking at keys.

6/8 Online restore is simpler: storing block-wide prefixes & timestamp suffixes separately allows quick updates without rewriting entire blocks—crucial for online restore which serves backups directly out of s3-like systems.

5/8 PrefixBytes ( is key: colbk groups 16 keys into a bundle, extracts a shared bundle prefix, and also stores a block-wide prefix. This means less repeated prefix decoding and no varint fuss.

4/8 Columnar blocks swap rows for columns, eliminating varints via a fixed layout (. The tricky part? Doing so without making blocks larger. Mission accomplished.

3/8 Historically, Pebble’s rowblk format (inherited from RocksDB) used row-oriented storage & varint encoding (space-efficient but adds decoding overhead). It also used prefix compression, storing full keys only every 16th row.

2/8 SSTables are sorted, immutable files of key-value pairs. They’re split into blocks, letting Pebble load only what’s needed from disk—crucial for CockroachDB’s performance.