Note: The talk is delivered by QuestDB's CTO and co-founder, Vlad, and Fast Data Advocate Javier.

For years, QuestDB ingested data via ILP (InfluxDB Line Protocol), a text-based, row-oriented format. We'd pushed it to ~8 million rows/sec. But at that scale, you're paying for overhead you shouldn't have to: UTF-8 number parsing on every value, row-by-row WAL appends, and repeated string-to-symbol-ID lookups. We designed a replacement from scratch and nearly doubled throughput to ~15 million rows/sec.
QWP (QuestDB Wire Protocol) is a binary, column-oriented ingestion protocol. Booleans are bit-packed, timestamps use Gorilla delta-of-delta compression (~1 bit per value for monotonic sequences), symbols use delta dictionaries, and the schema is hashed with XXH64 so repeated batches transmit 8 bytes instead of full column definitions.
We'll walk through the key design decisions:
Column-oriented wire format: the client groups values by column, so the server writes entire columns into the WAL via memcpy instead of row-by-row appends.
Zero-allocation WebSocket path: frame parsing operates directly on native memory buffers, masking/unmasking happens in-place, all cursors are reusable. No object allocation on the hot path.
Two-layer symbol caching: per-column and per-connection caches eliminate the repeated string lookups that dominated CPU time in the text protocol.
UDP for fire-and-forget: using Linux recvmmsg to batch-receive datagrams. Packet loss means silent data loss, and that's a documented, deliberate trade-off.
We'll show allocation profiling across ILP-over-TCP, ILP-over-HTTP, and QWP-over-WebSocket, and explain why a column-oriented wire format is a natural fit for a column-oriented storage engine. All code and the full wire specification are open source under Apache 2.0.