Goblin Core Documentation

Goblin Core is a compact, single-node server with Redis command semantics and two independent protocol choices: RESP for compatibility and SBE for typed binary clients. Both protocols work over TCP (optionally accelerated with Cisco ExaSock or native NVIDIA XLIO Ultra), Unix-domain sockets, shared-memory rings, polled one-sided RDMA rings, and Aeron UDP/IPC response channels. SBE also runs over provider-neutral libfabric reliable datagrams, including AWS EFA.

Start here

• Build, configure, and run Goblin Core
• Install Goblin Core from source
• Build and run the OCI image with Podman
• Release history
• Performance architecture
• Benchmark methodology and results
• BlueField DPU-side Pub/Sub benchmark
• Pub/Sub performance benchmark
• Libfabric provider matrix over 100 Gb/s
• XLIO Ultra command latency
• List storage algorithms
• Real-time and memory-efficient hash indexes
• Fixed-width packed sorted sets
• Firehose replication and Kafka recovery
• Kafka write-log ingestion
• Authentication and trusted transports
• TCP listeners and native TLS
• Hard memory ceiling and deterministic OOM

Protocols and transports

Protocol and transport are independent. RESP and opt-in SBE can each run over TCP (native TLS, ExaSock, or XLIO Ultra), Unix-domain sockets, a shared-memory ring, a polled RDMA ring, libfabric RDM, or Aeron UDP/IPC. XLIO preserves ordinary TCP on the wire and remains compatible with a kernel TCP peer. Ordinary socket TCP is always plaintext on 127.0.0.1 and requires TLS on non-loopback addresses. SBE requires --enable-sbe and is intentionally unauthenticated; see Authentication before exposing it.

Replication, durability, and replay

GOBLIN.FIREHOSE provides transport-neutral live replication over TCP, UDS, shared-memory rings, or RDMA. A replica can load a native snapshot, replay the durable RESP2 mutation log from Kafka, buffer the live firehose while it catches up, and open its listeners only after the handoff is complete. Kafka owns the durable history; Goblin Core remains the compact serving layer.

Use synchronous SAVE when the caller must know that the snapshot file is installed before continuing. Use forked BGSAVE to keep serving while an ordinary-page process writes a snapshot in the background. The command pages document accelerator formats, copy-on-write memory behavior, HugeTLB and transport restrictions, and the retained GOBLIN.* aliases.

The Thunder and Redpanda durability deployment record preserves the host-specific first production-style lab setup separately from the general source installation guide.

Command reference

Commands that share one implementation family link to the corresponding family reference. Goblin-specific atomic helpers and scripting commands have individual pages.

Server and keyspace

• AUTH, HELLO, CLIENT, COMMAND
• SELECT, QUIT, PING, ECHO
• INFO
• TIME, ROLE, CONFIG GET
• SAVE, BGSAVE, GOBLIN.LOAD
• MULTI, EXEC, DISCARD, WATCH, UNWATCH
• DEL, EXISTS, TYPE, DBSIZE, RANDOMKEY, TOUCH
• RENAME, RENAMENX, COPY, SCAN

The bounded-iteration reference documents the shared cursor, filtering, and mutation contract for SCAN, HSCAN, SSCAN, and ZSCAN.

The transaction reference documents atomic queued execution, optimistic WATCH, runtime versus queue-time errors, and the fixed page-backed per-client queue.

Strings and counters

• SET, SETNX, GET, GETEX, GETSET, GETDEL
• STRLEN, APPEND, GETRANGE, SETRANGE
• INCR, DECR, INCRBY, DECRBY, INCRBYFLOAT
• MSET, MSETNX, MGET

Expiration

• EXPIRE, PEXPIRE, EXPIREAT, PEXPIREAT
• TTL, PTTL, EXPIRETIME, PEXPIRETIME, PERSIST

Sorted sets

The sorted-set command reference covers conditional updates, score bounds, rank and score ranges, endpoint pops, scans, and storage. The packed sorted-set guide documents the six fixed-width member/score representations and their qualified command names.

• ZADD, ZINCRBY, ZCARD
• ZSCORE, ZMSCORE, ZRANK, ZREVRANK
• ZRANGE, ZRANGEBYSCORE, ZREVRANGEBYSCORE, ZCOUNT
• ZREM, ZREMRANGEBYRANK, ZINTERSTORE, ZUNIONSTORE, ZPOPMIN, ZPOPMAX, ZSCAN, ZREMRANGEBYSCORE

Hashes

The hash command reference covers behavior and replies. The hash implementation guide covers the efficient and incremental RT indexes, qualified command families, and --real-time mode.

• HSET, HMSET, HSETNX, HGET, HMGET
• HDEL, HEXISTS, HLEN, HSTRLEN
• HGETALL, HKEYS, HVALS, HINCRBY, HINCRBYFLOAT, HRANDFIELD, HSCAN

Bounded iteration

• SCAN, HSCAN, SSCAN, ZSCAN

Lists

The list command reference covers both the default segmented backend and the adaptive-PMA backend. The local work-queue benchmark compares the new commands with the existing list surface over one RESP2/UDS path.

• LPUSH, RPUSH, LPUSHX, RPUSHX
• LPOP, RPOP, LMOVE, RPOPLPUSH, LLEN, LINDEX, LRANGE
• BLPOP, BRPOP, BLMOVE, LMPOP, BLMPOP
• LSET, LTRIM, LREM, LINSERT

Every list command is also available under GOBLIN.PMA.* and GOBLIN.SEGMENTED.*, allowing clients and benchmarks to select a concrete implementation without changing the standard aliases.

Pub/Sub

The Pub/Sub command reference covers RESP2 arrays, RESP3 pushes, typed SBE delivery, glob matching, disconnect cleanup, and the page-backed slow-consumer queue.

• SUBSCRIBE, UNSUBSCRIBE, PSUBSCRIBE, PUNSUBSCRIBE
• PUBLISH, PUBSUB CHANNELS, PUBSUB NUMSUB, PUBSUB NUMPAT

Native atomic helpers

These commands replace multi-command Lua idioms with one atomic C++ operation.

Operational extensions are collected in the GOBLIN.* command reference: GOBLIN.MEMORY, GOBLIN.OPTIMIZE, GOBLIN.DUMPWORLD, and GOBLIN.LOAD. The legacy GOBLIN.SAVE and GOBLIN.BGSAVE spellings are exact aliases for the standard SAVE and BGSAVE commands. The same-version server stream is documented separately as GOBLIN.FIREHOSE.

Scripting

Each runtime has an independent VM and compiled-script cache. They share the same keyspace and atomic command execution model. See the scripting overview for argument, sandbox, cache, and return value rules.