EVAL

Executes a server-side Lua script.

Browse all commands · Upstream source documentation

Reference field	Value
Canonical name	`EVAL`
Catalog group	Scripting commands
Upstream since	2.6.0 — upstream Redis/module version, not an ffkv release
Pinned catalog	Redis 8.10.0

Syntax

EVAL script numkeys [key [key ...]] [arg [arg ...]]

ffkv behavior and limits

Sandboxed Lua 5.1 is bounded to 1 million instructions, 32 MiB Lua memory, 64 nesting levels, 8 MiB scripts/bridge values/replies, 10,000 bridge calls and 16 MiB cumulative bridge arguments. Writes before runtime errors commit. The volatile SHA cache holds 256 scripts/16 MiB with FIFO eviction; durable function libraries cap at 256 and one 16 MiB encoded WAL frame.

Source: README.md:426–449; tests/bdd/features/scripting.feature.

All requests share the 8 MiB framing/1,024-argument ingress limit; normal RESP replies have 64 MiB wire/logical budgets with bounded nesting/value counts. Management and Lua origins impose smaller budgets. These are not a process RSS cap. Catalog inclusion is not exhaustive option or wire-format parity. See limits and compatibility.

Upstream metadata

The following fields come from the pinned Redis metadata, not measured ffkv performance or an assertion that ffkv enforces every upstream flag. Running COMMAND INFO/COMMAND DOCS describe the current engine.

Upstream Redis complexity: Depends on the script that is executed.

Upstream arity: -3 (includes the command name; a negative number is a minimum).

Command flags: NOSCRIPT, SKIP_MONITOR, MAY_REPLICATE, NO_MANDATORY_KEYS, SCRIPT_RUNNER, STALE.

ACL categories: SCRIPTING.

Argument structure

Argument	Type	Constraints / metadata
`script`	string	Required
`numkeys`	integer	Required
`key`	key	optional; multiple; key_spec_index: 0
`arg`	string	optional; multiple

Key specifications

Upstream key discovery descriptors (indexes are relative to the full command vector):

[
  {
    "notes": "We cannot tell how the keys will be used so we assume the worst, RW and UPDATE",
    "flags": [
      "RW",
      "ACCESS",
      "UPDATE"
    ],
    "begin_search": {
      "index": {
        "pos": 2
      }
    },
    "find_keys": {
      "keynum": {
        "keynumidx": 0,
        "firstkey": 1,
        "step": 1
      }
    }
  }
]

Additional pinned upstream metadata

{
  "group": "scripting",
  "since": "2.6.0"
}

Metadata source · Local snapshot: src/command_metadata.json, key eval.

BDD coverage and source examples

These are source observations, not a test run performed by this documentation build. A tag identifies scenario intent; an extracted vector identifies a command invocation, which may intentionally produce an error. Scenarios may require setup, multiple connections, persistence, cluster nodes or native-module fixtures.

Lua preserves arbitrary decimal replies and sparse cardinality beyond signed integers (Scenario Outline) — tests/bdd/features/arrays.feature:434; command tag and invocation observed.
Disposable debugging restores an absent cross-database destination (Scenario Outline) — tests/bdd/features/client_control_repair.feature:247; invocation observed.
A disconnected debugger queued behind a native command cannot strand the actor (Scenario Outline) — tests/bdd/features/client_control_repair.feature:300; invocation observed.
Nested Lua key denial retains the calling identity and Lua context — tests/bdd/features/client_control_repair.feature:380; command tag and invocation observed.
Dynamic index namespace denials are logged without logging visibility filtering — tests/bdd/features/client_control_repair.feature:474; command tag and invocation observed.
Dynamic Lua and function keys cannot diverge from their authoritative cluster owner — tests/bdd/features/cluster.feature:1279; command tag and invocation observed.
Dynamic script admission shares importing and migrating key visibility — tests/bdd/features/cluster.feature:1379; command tag and invocation observed.
Scripts use invocation-local layouts rather than the network session — tests/bdd/features/hashes.feature:1253; invocation observed.
A script sees its own writes and earlier effects survive a later Lua error — tests/bdd/features/scripting.feature:4; command tag and invocation observed.
A line debugger pauses a live script and kill interrupts its eligible execution — tests/bdd/features/scripting.feature:377; invocation observed.
Debug stepping exposes the first write and prevents unsafe cancellation — tests/bdd/features/scripting.feature:418; invocation observed.
Fork style debug execution discards effects while retaining a useful result — tests/bdd/features/scripting.feature:471; invocation observed.
A no-writes declaration is enforced rather than trusted — tests/bdd/features/scripting.feature:597; command tag and invocation observed.
Each outer Lua invocation owns a fresh import session rather than borrowing network layouts — tests/bdd/features/scripting.feature:623; command tag and invocation observed.
Debug mode validates its argument and NO restores ordinary execution — tests/bdd/features/scripting.feature:719; invocation observed.
A valid write-capable shebang script rejects readonly entry points — tests/bdd/features/scripting.feature:770; command tag and invocation observed.
Disposable Lua debugging sees staged search changes without publishing catalog changes — tests/bdd/features/search.feature:1316; invocation observed.
Native commands run inside transactions and scripts unless their registered flags forbid it — tests/bdd/features/server.feature:917; invocation observed.
Native command metadata exposes actual arity flags and key positions — tests/bdd/features/server.feature:946; invocation observed.
Invalid counted script keys fail during EXEC without discarding adjacent writes (Scenario Outline) — tests/bdd/features/transactions.feature:289; command tag and invocation observed.
Invalid script arity still poisons the transaction at queue admission — tests/bdd/features/transactions.feature:337; command tag and invocation observed.
Queue admission checks command ACLs even when script key counts are malformed — tests/bdd/features/transactions_runtime.feature:271; command tag and invocation observed.
Script key ACLs are checked at queue admission and rechecked during execution — tests/bdd/features/transactions_runtime.feature:314; command tag and invocation observed.

Extracted command vectors

The following JSON arrays are the test harness’s argument vectors, not shell commands. Numeric arguments are converted by the harness; objects of the form {"base64":"…"} represent decoded binary bytes, not a literal JSON command argument. Preserve scenario setup and connection ordering. See the CLI guide for shell quoting and binary stdin. No reply is invented or inferred.

A disconnected debugger queued behind a native command cannot strand the actor — tests/bdd/features/client_control_repair.feature:330 (client “new-debugger” executes:)

Expanded from an actual Examples row: script=return redis.call('SET','abandoned','must-not-exist').

[
  "EVAL",
  "return 42",
  0
]

Next source assertion: the reply equals:

A line debugger pauses a live script and kill interrupts its eligible execution — tests/bdd/features/scripting.feature:408 (I execute:)

[
  "EVAL",
  "return 7",
  "0"
]

Next source assertion: the reply equals:

Invalid counted script keys fail during EXEC without discarding adjacent writes — tests/bdd/features/transactions.feature:295 (I execute:)

Expanded from an actual Examples row: command=EVAL, count=1, error=ERR Number of keys can't be greater than number of args.

[
  "EVAL",
  "return 1",
  "1"
]

Next source assertion: the reply equals:

"QUEUED"