Environment variables

500

A usize

The maximum number of keys that should be scanned at once in general queries. Read only before 3.1.0; later versions use a fixed batch of 500.

1000

A usize

The maximum number of keys that should be scanned at once for export queries.

50,000 (10,000 on 2.x)

A usize

The maximum number of keys that should be scanned at once for count queries. Read only before 3.1.0; later versions use a fixed batch of 50,000.

250

A usize

The maximum number of keys to scan at once per concurrent indexing batch. Read only before 3.1.0; later versions use a fixed batch of 250.

1000

A u32

The number of document ids allocated at once from a table's shared document id sequence, which full-text, HNSW and DiskANN indexes use. Larger batches reduce coordination on the distributed sequence, at the cost of larger gaps in the ids when a node is lost before it uses up its current batch. Replaces

, which 3.3.0 no longer reads.

1024

A usize

How many graph adjacency keys one background fold pass packs into blocks per transaction. This bounds the transaction size on a vertex with many edges, which is then folded over several passes. See

.

512 (10,000 before 3.1.0)

A usize

Specifies the number of items which can be cached within a single transaction.

1,000

A usize

The number of definitions which can be cached across transactions.

268,435,456 (256 MiB)

A usize

The maximum total size, in bytes, of the HNSW vector cache, shared across all HNSW indexes in the process. This bounds the cached element

used for distance computation - it is not a cap on total HNSW memory: the adjacency graph is loaded into resident memory on first use and stays there, outside this budget. Contrast

, which does page graph structure.

268,435,456 (256 MiB)

A usize

The maximum total size, in bytes, of the DISKANN index cache, shared across all DISKANN indexes in the process. DISKANN graph data lives in the key-value store and is paged through this bounded cache. The DISKANN counterpart of

.

134,217,728 (128 MiB)

A u64

The maximum total size, in bytes, of the cache of compacted full-text terms, which every read-only full-text query on the node shares. Changes that are not compacted yet are always read from storage, and write transactions do not use the cache.

turns it off.

67,108,864 (64 MiB)

A u64

The maximum total size, in bytes, of the cache that maps document ids back to record keys when a traversal reads numeric-encoded graph adjacency entries. Only relevant once

has written numeric blocks.

none (deny all)

Colon-separated paths on Unix; semicolon-separated on Windows

Directories the server may read when an analyzer uses

. An empty or unset value denies every path. Each mapper file must resolve under one of the listed directories. See

.

none (use the dictionaries built into the binary)

A path to a directory

Directory holding the morphological dictionaries the

tokenizer reads, one subdirectory per language:

,

,

. The released binaries already carry all three, so this is only needed to replace them with a custom or updated dictionary, or for a build made without them. Each language is overridden only if the directory holds a dictionary for it; the rest keep using the built-in one. See

.

none

Comma-separated paths

Specifies a list of paths in which files can be accessed.

none

A string

The backend URL of the global bucket, in the same form as the

of a bucket, such as

,

or an object storage URL. A bucket uses part of it through a

backend. See

.

false

true, false

Whether every bucket must use the global bucket. When

,

is the only backend a bucket can have, and only a root user can give a bucket that backend.

none (presigned URLs off)

A string of at least 32 bytes

The secret that signs

. Without it, or with a shorter value,

and

fail and a signed request is refused. Set the same value on every node of a cluster. Changing it revokes every URL already issued.

1h

A duration, up to 7d

The longest

a presigned URL can have. A larger value is reduced to 7 days.

5 GiB

A byte size, such as

The largest

a presigned upload URL can have, and the

of one created without it.

none

An origin with a scheme and no path, such as

The address at which clients reach the server, returned as the

of a presigned URL and placed before its

in

. A value without a scheme or with a path is ignored, and

is then

.

10

A usize

The maximum number of HTTP redirects allowed within http functions.

128

A usize

The maximum number of idle HTTP connections to maintain per host.

1000

A usize

The maximum number of idle HTTP connections to maintain. Has no effect: earlier versions read the value without applying it to the HTTP client, and 3.3.0 no longer reads it. Use

instead.

90

A usize

The timeout for idle HTTP connections before closing.

30

A usize

The timeout for connecting to HTTP endpoints.

SurrealDB

A string

The USER-AGENT string used by HTTP requests. Before 3.1.0 this variable was named

, which later versions ignore.

1,048,576

A usize

How many concurrent network requests can be handled at once

4,194,304 (4 MiB)

A usize

Maximum HTTP body size of the HTTP

endpoint. See Embedded MCP.

1,048,576 (1 MiB)

A usize

Maximum HTTP body size of the HTTP /sql endpoint

1,048,576 (1 MiB)

A usize

Maximum HTTP body size of the HTTP

endpoint. See

.

1,048,576 (1 MiB)

A usize

Maximum HTTP body size of the HTTP

endpoint. See

.

4,194,304 (4 MiB)

A usize

The maximum HTTP body size of the HTTP /api endpoint.

5,368,709,120 (5 GiB)

A usize

The maximum size of an upload to the HTTP

route, whether it uses a session or a presigned URL.

4,194,304 (4 MiB)

A usize

Maximum HTTP body size of the HTTP /rpc endpoint.

16384

A usize

The maximum number of attached HTTP RPC sessions that a node keeps. A session counts against the node it was attached through until it is detached or expires, wherever it is used. When the limit is reached, the node first releases the sessions that were detached or expired through other nodes, and then refuses the attach.

1024

A usize

How many notifications can be buffered for one GraphQL subscription. When the buffer is full, new notifications for that subscription are dropped.

The value of

, so 4,194,304 (4 MiB) unless that is raised

A usize, clamped to a ceiling of 4,294,967,295 (4 GiB less one byte), the largest size a gRPC length prefix can express. A value below the floor needed to carry one streaming file chunk is raised to it.

From 3.3.0. The largest single gRPC message the server will accept or emit, also reported to clients as

. This is the gRPC counterpart of

: raise it for a transaction that writes a large batch in one request.

16384

A usize

The maximum number of explicitly attached sessions the gRPC transport keeps at once. The count is for the whole process rather than per connection, because a gRPC channel reconnects transparently and one session's requests can arrive on different connections. Per-request sessions do not count towards it.

1024

A usize

How many live query notifications can be buffered for a gRPC subscription whose client is not keeping up. When the buffer is full, the server ends that subscription with a

status instead of waiting.

16,384 (16 KiB)

A usize

Maximum HTTP body size of the HTTP /key endpoints

1024 (1 KiB)

A usize

Maximum HTTP body size of the HTTP /signup endpoint.

4,294,967,296 (4 GiB)

A usize

Maximum HTTP body size of the HTTP

endpoint. Counted cumulatively across the whole request, not per chunk. Also enforced on the gRPC import stream. See Request size limits.

1024 (1 KiB)

A usize

The maximum HTTP body size of the HTTP /signin endpoints

60

Seconds (integer);

disables

Outer timeout on each MCP tool execution.

262,144 (256 KiB)

Bytes (integer);

disables

Maximum serialised tool / resource response size; larger results are truncated with a marker.

64

A positive integer

Maximum arguments for a single

tool call.

256

A positive integer

Maximum top-level keys in MCP parameter / data objects.

4,096 (4 KiB)

A positive integer; values ≤ 0 fall back to the default

Maximum byte length of a single

SurrealQL pass-through string inside a

payload.

200

A positive integer; values ≤ 0 fall back to the default

Maximum tables the database-level schema resource enriches with per-table fields / indexes / events. Tables beyond the cap keep a bare

and the body includes a

marker.

loopback only (

,

,

)

Comma-separated exact hostnames

Hostnames accepted in the HTTP

header for

(DNS-rebinding guard). A non-empty list

the loopback default, so include

yourself if you still need it. Entries without a port match any port. Ignored when

is set.

/

to enable

Disables the

-header allowlist (accept any

). Escape hatch for a trusted proxy or load balancer. Takes precedence over

.

1,000,000

A usize

Per source row, bounds live and emitted rows during variable-length path expansion in a GQL

.

1,000,000

A usize

Bounds the in-memory build side of hash joins between GQL match patterns (and distinct-operator working sets).

1,000,000

A usize

Bounds cumulative rows emitted by GQL hash-join operators (including cross joins).

64

A number of bytes, without a unit suffix

Largest encoded payload an

field set may occupy inside a graph adjacency entry. A larger payload spills, leaving the values on the edge record only.

250,000

A usize;

disables the budget

How many index entries one range branch of a bitmap plan, which combines several indexes into a bitmap of candidate records, may read before that branch is abandoned. The abandoned branch's condition is then checked as an ordinary

filter, so an unselective range cannot make the bitmap plan slower than scanning. The same budget is the upper limit for the allow-list built for a pre-filtered vector search, which from 4.0 has a smaller limit when the search width is small (see

).

50000

A usize

The number of result records which will trigger on-disk sorting.

16384

A usize

The maximum input string length for similarity/distance functions

4

A usize

A batch with fewer items than this is evaluated one item at a time. A larger batch has its items evaluated concurrently, but only when the evaluation is read-only. Set it above the batch size to turn concurrent evaluation off.

20 (2

= 1,048,576 bytes)

A power of two, as its exponent: an integer up to 28 (256 MiB); larger values are treated as 28

Limits the memory that a single call to

,

,

,

,

,

,

,

,

,

or

can allocate for its result, and the number of values that a scripting function can return. A call over the limit fails with an error. The value is an exponent, not a byte count: the limit is 2 to the power of the value, so the default of

allows 1,048,576 bytes. Setting a byte count such as

sets the limit to its maximum of 2

bytes.

256

A usize

The maximum recursive idiom path depth allowed. Read only before 3.1.0; later versions use a fixed limit of 256.

0 (half of

if set, otherwise 2 GiB)

A usize or suffixed integer

The bytes that all queries on the node may hold in sorts, grouped aggregates,

and join builds before they spill to disk. See

.

0 (the node memory limit)

A usize or suffixed integer

The bytes that a single query may hold which it cannot spill, such as gathered values and its transaction buffers. A query that passes it fails with the memory threshold error, and other queries are not affected.

4 MiB

A usize or suffixed integer

The bytes that every query holding a sort, group,

or join build is guaranteed, in addition to what spilling takes.

1000

A number of milliseconds

The longest that a query waits in total for other queries to release memory before it continues past the node limit.

never waits.

1 MiB

A suffixed integer greater than zero

The size of the parts in which a spilled run is written and read.

8

A usize, at least 2

The number of spilled runs that a merge reads at once. A merge of more runs takes several passes.

256 KiB

A suffixed integer greater than zero

The bytes that a sort, group,

or join build reserves from its query at a time.

No limit

A byte count, or a size such as

The estimated bytes of result rows that a write statement streaming its results, in a transaction of its own, holds back until it commits. Held rows let a statement that meets a retryable conflict at commit run again without its client receiving the rows of the failed attempt. Past this size the held rows are sent and the rest stream as they are produced, and the statement is no longer run again on a conflict.

holds nothing back. By default every row is held until the commit.

120

A usize

Specifies how deep recursive computation will go before erroring.

64

A usize

Specifies how many concurrent jobs can be buffered in the worker channel. Removed in 4.0.0, which no longer reads it, so setting it has no effect.

128

A usize

How deep the parser will build an expression operator tree. This bounds long operator chains such as

and prefix/postfix chains, which consume neither of the other two parsing budgets.

100

A usize

How deep the parser will parse nested objects and arrays. Type annotations such as

count against it as well. For data arriving over an SDK or RPC this is the governing limit. For a query written as text, see

.

1000

A usize

The maximum size of the priority queue triggering usage of the priority queue for the result collector.

20

A usize

How deep the parser will parse recursive queries (queries within queries). Object and array literals written into a query count against this budget as well as against

, and this one is the smaller of the two - see

.

2

A usize

The number of batches each operator buffers ahead of downstream demand. Set to 0 to disable operator-level pipeline buffering.

10,485,760 (10 MiB)

A usize

Limits the maximum allowed size (in bytes) for regular expressions. This prevents excessive memory consumption when building complex or very large regex patterns.

2,097,152 (2 MiB)

A byte count, or a size such as

The approximate decoded size of each batch of records that a query fetches from an index result, such as the matches of a full-text search. A batch also holds at most 1,000 records. Every operator after the scan holds whole batches, so the memory a query uses grows with this value. Lower it to limit memory on a small node, or raise it for fewer, larger batches.

and values that cannot be parsed are ignored.

0 (disabled)

A

;

disables

Maximum number of key writes a single statement transaction may buffer before it is aborted and rolled back. Cascaded deletes, full-text index maintenance, graph-edge cleanup, and commit-time changefeed / live-query events all count toward the limit; each range delete counts as one write. Tripping the guard poisons the transaction so an explicit client

cannot persist a partial statement. Internal maintenance transactions (index builds, compaction, GC) are not bounded. On TiKV, range deletes may expand further inside the storage layer - size distributed clusters with both this limit and

in mind. Scale / multi-node deployments often use a limit in the 50,000-100,000 range; leave at

for typical single-node use.

3600 (one hour)

A u64 (seconds)

The largest

that a

can ask for. A larger request fails when the transaction begins, with

. It also limits how long a permission that has been revoked can still apply to a stale read.

allow

,

,

What a read-only transaction with

does when no published snapshot is recent enough: read the current state (

), fail (

), or read the current state at most

times a second.

on the statement fails in that case whatever this setting says. A datastore that publishes no snapshots, which includes every storage engine of the open-source release, always reads the current state, which meets every bound.

Number of CPU cores (minimum 16)

A usize, at least 4

Number of worker threads the in-memory and SurrealKV engines use for storage work. A value below 4 or a value that is not a number logs a warning and the default is used.

512

A usize

Number of threads which can be started for blocking operations.

10,485,760 (10 MiB)

A usize

Runtime thread memory stack size. Default stack size is doubled if compiled from source in Debug mode.

Number of CPU cores (minimum 4)

A usize

Number of runtime worker threads used to start. See

.

No limit

A usize

The most places of the node's worker pool that one statement can hold at once to scan in parallel. The pool has one place for each core beyond the one a statement's own task runs on.

runs every statement on its own task. Index builds and background index compaction draw on the same pool.

1

A usize

The places of the worker pool that parallel scans never take, so that short statements keep a core. On a host with two cores or fewer, the default leaves no place to share, so index builds and compaction run on one task unless this is set to

.

8

A u32

How many chunks of its work a parallel worker finishes while another statement waits for a place, before it hands what is left of its work back and frees the place.

50000

A u32

The most records in a million that a scan's

conditions can keep for the scan to be split into parts that run in parallel. The default is one record in twenty. A scan that keeps a larger share of the records it reads stays on one task, because most of its time is spent reading records, which more workers do not make faster. The share is measured from the records the scan has read so far. With

, a scan with a

condition can be split whatever share of its records the condition keeps.

262_144 (256 KiB)

A usize

Maximum stack size of the JavaScript function runtime.

2,097,152 (2 MiB)

A usize

Maximum memory limit of the JavaScript function runtime.

5000 (5000 milliseconds or 5 seconds)

A usize

Maximum allowed time in milliseconds that a JavaScript function is allowed to run for.

false

true, false

Forward all signup/signin/authenticate query errors to a client performing authentication. Do not use in production.

300

seconds

The shortest time between two fetches of a

object that a token with an unknown key identifier (

) can cause.

43200

seconds

How long a fetched JWKS object stays in the cache before it is fetched again. The default is 12 hours.

1000

milliseconds

How long a request for a JWKS object from its URL may take before it fails. Raise it for an identity provider that answers slowly.

false

true, false

Whether to lazy-load Surrealism modules instead of eagerly compiling them at server startup.

100

A usize

The number of surrealism modules which can be cached across transactions.

8

A usize

Per-module controller pool size ceiling for Surrealism WASM modules. Has no effect: the server does not read this variable, so the ceiling stays at 8. Removed in 4.0.0, which runs each call on a fresh instance; see

.

None (unlimited)

A usize

Per-module WASM linear memory ceiling in bytes. From 4.0.0, an extension's

can set a lower limit, and the smaller of the two applies.

None (unlimited)

A u64 (milliseconds)

Per-invocation execution time ceiling for Surrealism WASM modules. From 4.0.0, an extension's

and the query's timeout can set a lower limit, and the smallest applies.

64

A u32

How many extension calls can run at once across the datastore. Each call runs on a fresh instance in one of these slots, and each slot reserves about 4 GiB of virtual address space (not resident memory) on 64-bit hosts, which a virtual memory limit such as

must allow. A quarter of the slots is kept for calls made from inside another extension. A call that finds no free slot waits, within its query's timeout.

104857600 (100 MiB)

A u64

The largest total size, in bytes, of the files that an extension archive bundles in its read-only filesystem. Applied when an archive is unpacked by

or at startup.

false

true, false

Whether an extension defined with

runs without time checks. Without this setting, every call is stopped by the execution time ceiling and the query's timeout, whatever the definition says. With it, such an extension holds its slot and its worker thread until it returns.

None (unlimited)

A usize

Per-module KV store entry count ceiling. Has no effect from 4.0.0, which removes the module KV store.

None (unlimited)

A usize

Per-module KV store maximum value size in bytes. Has no effect from 4.0.0, which removes the module KV store.

debug

none, full, error, warn, info, debug, trace

Controls the tracing level at which Surrealism module stdout is emitted.

A URL

Not read from 4.0.0, where

packages always come from

. Before 4.0.0: the base URL that a

package in

is fetched from, as

. A trailing slash is removed. Set it to a mirror to serve packages from somewhere else.

true

true, false

Whether the Prometheus

endpoint is served on the HTTP port. When

, the route returns

. See

.

5

Seconds (integer); values below 1 are raised to 1

How often the process memory and CPU snapshot is refreshed. The

gauges and

read this snapshot.

1000

Milliseconds (integer);

disables the counter

The duration above which a completed statement is also counted by the

metric.

false

true, false

Whether to disable sending metrics to the GRPC OTEL collector.

false

true, false

Whether to disable sending traces to the GRPC OTEL collector.

none

A string

If set then use this as value for the namespace label when sending telemetry

none

otlp

If set to "otlp" then telemetry is sent to the GRPC OTEL collector.

false

true, false

Whether to enable tokio console.

60

A usize

How long, in seconds, to retain data for completed events.

none

String to a socket address

The socket address that Tokio Console will bind to.

auto

,

, or a number of tasks (

means

)

How many tasks one batch of HNSW index compaction builds its graph changes on.

uses as many tasks as the machine reports available parallelism, and

uses one. A batch that removes vectors always runs on one task.

auto

,

, or a number of tasks (

means

)

How many batches of records the initial scan of a standard or

index build indexes at the same time, each in a transaction of its own.

uses the machine's available parallelism, up to 16, and

indexes one batch at a time.

64

A usize greater than zero

The first and smallest number of candidate records that a filtered KNN search on a DiskANN index reads at once. Each following read doubles, up to

.

4096

A usize greater than zero

The largest number of candidate records that a filtered KNN search on a DiskANN index reads at once.

true

true, false

Whether KNN queries build an allow-list from their

conditions. When

, the conditions are checked for each candidate during the index traversal instead, as before 3.3.0.

2000

A u64

The allow-list size at or below which the search does not use the index graph at all, and instead scores every allowed record exactly. At this size an exact search is cheaper than a traversal and always finds the true nearest neighbours.

5

A u32

How strongly an HNSW search above

favours walking the graph over scoring the allowed records exactly. The search scores them exactly when the number of allowed records, squared and multiplied by this value, is at most

multiplied by the index's

and by the number of nodes in the graph. A larger value keeps more searches on the graph.

4

A u32

The multiplier applied to the search width (

) of a DiskANN traversal restricted to an allow-list. From 4.0, an HNSW traversal reads

instead. A selective allow-list rejects most candidates, and a wider search recovers the recall that loses. The result is capped by

and is never lower than the

the query asks for.

1

A u32

The multiplier applied to the search width (

) of an HNSW traversal restricted to an allow-list. The traversal already continues past the candidates the allow-list rejects, so the default of

does not widen it. Set it to

for the width that 3.x used.

100,000

A u64

The allow-list size above which the traversal stops multiplying

, because a large allow-list rejects few candidates. Results are still restricted to the allow-list at every size.

1024

A u32

The ceiling for the multiplied

. It never lowers the

a query asks for.

true

true, false

Whether a KNN search on a DiskANN index of

or

compares candidates using an 8-bit code of each vector, and then ranks its nearest results from the full vectors. The distances a query returns are exact either way.

compares full vectors throughout. Only an index defined or rebuilt on SurrealDB 4.0 keeps the codes, and other indexes ignore the setting.

1

A usize

How many candidates a DiskANN search expands in each step. A wider beam reads the neighbours of all of them in one batched read, so the search makes fewer storage round trips, but it also expands some candidates that a narrower beam would skip. The value is limited to between 1 and the search size of the query.

134,217,728 (128 MiB)

A usize

The maximum WebSocket message size, which is also used as the maximum frame size. From 3.0.0 this replaces

, which no longer exists and has no effect if set.

Greatest possible usize value

A usize

The maximum write buffer size before backpressure is applied.

131,072 (128 KiB)

A usize

The size of the read buffer for WebSocket connections. This controls how much data can be buffered when reading from WebSocket connections. Larger values can improve performance for high-throughput connections but consume more memory per connection.

131,072 (128 KiB)

A usize

The size of the write buffer for WebSocket connections. This controls how much data can be buffered when writing to WebSocket connections. Larger values can improve performance for high-throughput connections but consume more memory per connection.

0

A usize

How many responses can be buffered when delivering to the client.

100

A usize

Number of messages that can be queued for sending via WebSocket. From 3.3.0, replies and live query notifications have separate queues on each connection, each holding this many messages, and a value below 2 is raised to 2. A connection whose notification queue is full is closed, with a close frame giving the reason, rather than dropping notifications.

3

A usize

How often (in milliseconds) any buffered responses are flushed to the WebSocket client.

32

A usize

Maximum number of streaming queries (the

RPC method) running at once on a single WebSocket connection. Each open stream holds an executing query, with its read snapshot or transaction, for as long as the client keeps reading. A request over the limit is refused with an error rather than queued.

30

Seconds (integer), clamped to 1-86,400

How long a streaming query may wait to hand one frame to the connection before the stream is stopped. The wait resets for every frame, so it limits a stalled client rather than the length of a stream, and it applies even when

is not set.

256

A usize

The maximum number of explicitly attached sessions that a single WebSocket connection can hold at once. The sessions of a connection end when it closes.

64

A usize

Maximum concurrently open client-managed transactions (

RPC) on a WebSocket connection's implicit default session. Further

calls return

until a slot is freed by

,

,

, or disconnect. Does not apply to attached sessions (see

).

64

A usize

Maximum concurrently open client-managed transactions within a single attached WebSocket session. Counted per session and independent of the connection default-session limit. Detaching or resetting a session cancels its open transactions and frees their slots.

none (non-deterministic)

A

Seeds the engine-wide RNG for reproducible runs in benchmarks and tests. Do not use in production.

,

Selects the live-query execution engine.

is the historical behaviour;

decouples write cost from subscriber count (experimental rollout).

1h

A duration

How long the router engine retains live-query event history for subscriber resume. Only applies when

.

none

A

Seeds HNSW index construction for reproducible graph layouts in benchmarks. Do not use in production.

false

true, false

Whether background graph folds write numeric-encoded adjacency blocks, which store each entry's record id as a compact table-level document id. Existing blocks are converted on their next fold, and reads handle both encodings. Size the lookup cache with

.

Automatically populated

A string

The version identifier of this build. Defaults to the CARGO_PKG_VERSION environment variable if not specified.

never

never|sync|async

Append-only log mode. Only used by the memory engine.

none

String to a file path

Filesystem path for persistence. Only used by the memory engine.

0 (unlimited)

A duration

Version retention period as a duration string. Used by the surrealkv engine. From 3.3.0 the memory engine refuses a non-zero value and the server does not start.

never

never|duration

Snapshot interval. Only used by the memory engine.

every (never on the memory engine)

never|every|duration

The sync mode for the database, the same setting as the

path parameter. Used by the memory and surrealkv engines. When both are set, it takes precedence over

.

every (never on the memory engine)

never|every|duration

The sync mode for the database. Used by the memory and surrealkv engines.

false

true|false|0|1

Whether MVCC versioning is enabled, for use with the

. Used by the surrealkv engine. From 3.3.0 the memory engine does not support versioned queries, and refuses

so that the server does not start.

false

true|false|0|1

Lets a 4.0 node open a datastore that a 3.x release wrote, and upgrade it in place. The upgrade is experimental. Without this variable, the node refuses to start on such a datastore and leaves the data unchanged. A datastore that 4.0 has already upgraded opens without it. An embedded SDK reads the

datastore option instead. See

.

none

4.0|current

The record format release a datastore is switched to when a node starts. A datastore in memory or SurrealKV switches to the 4.0 format on its first 4.0 start without it. A TiKV datastore keeps the 3.x format until a node starts with this set, and the switch is refused, with a warning, while a node that runs an older release is still running. It never moves a datastore back to an earlier format. See

.

0 (disabled)

A usize or suffixed integer

The tracked memory, in bytes, above which queries are refused with

. The value can be specified as bytes (b, or without any suffix), kibibytes (k, kb, or kib), mebibytes (m, mb, or mib), or gibibytes (g, gb, or gib); Kubernetes-style

and

are not accepted. Unset or

disables the guard, and any other value is raised to at least 1 MiB. Tracked memory is the size the allocator reserves for each live heap allocation, which includes the SurrealKV block cache and is less than the process's resident memory, so set the threshold below the container's memory limit and above

: at or below the cache size, a full cache refuses every query until the server restarts. Before 4.0, the RocksDB block cache was counted in the same way, and the threshold had to be above

. From 3.3.0 the reserved size rather than the requested size is counted, so a threshold carried over from an earlier release refuses queries at a lighter workload and may need raising.

From 3.3.1 the threshold also stops a query that is already running: a query that reads records checks it after each batch, and fails with the same error once tracked memory is above it, rather than growing until the operating system stops the server.

From 4.0.0 it also sets the default node memory limit, which is half of the threshold when

is not set. The threshold still refuses queries when the whole process passes it, whether or not their sorts and groups would spill. See

.

60

Seconds (integer);

disables the check

How long the memory threshold may refuse requests without a break before

answers

, so that a liveness probe restarts the process. Any observation below the threshold resets the clock. Has no effect while

is not set.

a

folder in the current directory

A URL, such as

Where SurrealML model files were stored. SurrealDB 4.0 removed SurrealML, so a 4.0 server stores nothing here. The folder is created when it does not exist.

a

folder in the current directory

A URL, such as

Where SurrealML model files were cached. SurrealDB 4.0 removed SurrealML, so a 4.0 server caches nothing here. The folder is created when it does not exist.

1000

An integer from 10 to 100,000; a value outside that range is treated as the nearest end of it

The number of computed regexes which can be cached in the engine.

start

All origins

A comma-separated list of origins

The origins allowed to make cross-origin (CORS) requests, such as

. When it is not set, every origin is allowed. When it lists origins and no

, a WebSocket connection to

that sends an

header is accepted only from a listed origin or with

.

start

5s

A duration

The interval at which to process async events.

export, import, sql

root

root, namespace, ns, database, db

Authentication level to use when connecting.

start

127.0.0.1:8000

String to an address

The hostname or IP address(es) to listen for connections on.

start

(disabled)

String to an address

The hostname or IP address to listen for

connections on.

start

(disabled)

String to an address

The hostname or IP address to listen for

connections on, for example

. The listener is disabled when this is not set.

start

false

true, false

Allow all capabilities.

start

none

guest, record, system (comma-separated)

Allows arbitrary queries to be used by user groups except when specifically denied. Alternatively, you can provide a comma-separated list of user groups to allow specifically denied user groups to prevail over any other allowed user group.

start, sql

none (denied for all subjects)

guest, record, system (comma-separated)

Allow

and

for listed subject groups. Not enabled by

. Still subject to

restrictions -

blocks

for that subject even when eval is allowed here. For remote clients, set on the

process only. See

.

start, sql

none

surrealism, kvexport (comma-separated)

Allow execution of experimental features. For remote clients, set on the

process. On

, affects embedded engines and REPL parse validation only. See

.

turns on the

route. From 3.3.0, ISO GQL is on by default and does not need this variable, and from 4.0 the same applies to buckets and files. From 4.0, the values

and

are refused as invalid names, and the server does not start.

start

all functions allowed

Empty,

, or comma-separated function paths

Allow execution of all functions except for functions that are specifically denied. Set to an empty value or

to allow all functions. Use a comma-separated list (for example,

) to allow specific function families or names. Values such as

are not valid. The environment variable name is singular (

), matching the

flag.

start

false

true, false

Allow guest users to execute queries.

start

all routes allowed

Empty, or a comma-separated list of HTTP routes

Allow all HTTP routes except those specifically denied, or only the listed routes. Route names:

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

,

.

start

all tools allowed

Empty, or a comma-separated list of MCP tool names

Allow all tools of the built-in MCP server except those specifically denied, or only the listed tools.

start

none

Comma-separated list of paths

Allow all or certain outbound network access.

start

all methods allowed

Empty, or a comma-separated list of RPC methods

Allow all RPC methods except those specifically denied, or only the listed methods.

start

false

true, false

Allow execution of embedded scripting functions.

start

false

true, false

Takes a boolean. From 2.5.0, closures cannot be stored, which removes a potential attack surface; on 2.5.x and 2.6.x this capability allows them again. Removed in 3.0.0, where closures can never be stored and the flag is rejected.

start

false

true, false

Deny all capabilities.

start

none

guest, record, system (comma-separated)

Deny arbitrary queries for the listed user groups, or for all of them when no value is given. Specifically allowed user groups prevail over a general denial.

start, sql

none

guest, record, system (comma-separated)

Deny

and

for listed subject groups. Deny prevails over allow at the same specificity. For remote clients, set on the

process.

start

false

true, false, comma-separated list

Deny execution of all or certain functions.

start

true

true, false

Deny guest users from executing queries.

start

none

Empty, or a comma-separated list of HTTP routes

Deny all HTTP routes except those specifically allowed, or only the listed routes. Uses the same route names as

.

start

none

Empty, or a comma-separated list of MCP tool names

Deny all tools of the built-in MCP server except those specifically allowed, or only the listed tools.

start

true

true, false, comma-separated list

Deny all or certain outbound access paths.

start

none

Empty, or a comma-separated list of RPC methods

Deny all RPC methods except those specifically allowed, or only the listed methods.

start

true

true, false

Deny execution of embedded scripting functions.

start

false

true, false

Issue the access token as an

cookie through

and

, and accept the cookie from the origins in

. When the record access method has

, the refresh token is issued as a second

cookie, and

rotates both. Requires

with no

entry. See

.

start

30s

A duration

The interval at which to perform changefeed garbage collection.

start

socket

none, socket, CF-Connecting-IP, Fly-Client-IP, True-Client-IP, X-Real-IP, X-Forwarded-For, Forwarded

The method of detecting the client's IP address.

parses the RFC 7239

header (

parameter).

sql

main

A string

The database selected when starting the REPL.

export, import

none

A string

The database selected for the import or export.

start

main

A string

The default database to use when starting a SurrealDB instance.

start

main

A string

The default namespace to use when starting a SurrealDB instance.

start

false

true, false

From 4.0.0 this variable has no effect: every attached HTTP RPC session is stored in the datastore, and the server still accepts the flag and the variable, and logs a deprecation warning. Before 4.0.0 it persists client-attached HTTP RPC sessions in the datastore so they survive restarts and can be resumed on any node sharing the storage, and is off by default. WebSocket sessions are not durable. The persisted copy includes authentication state and is stored unencrypted. See

.

start

1h (60s before 4.0.0)

A duration;

disables the background sweep

How often expired attached HTTP RPC sessions are purged from the datastore. An expired session is refused whether or not it has been purged, so the sweep only reclaims storage. Before 4.0.0 it applies only when durable sessions are enabled.

start

24h

A duration of at least 1ms

Idle lifetime of an attached HTTP RPC session; each use refreshes the expiry (sliding TTL). From 4.0.0 a session expires between one and one and a half times this value after its last request, and applies to every attached HTTP RPC session. Before 4.0.0 it applies only when durable sessions are enabled.

start

5s

A duration

The interval between background passes that fold a vertex's per-edge graph adjacency keys into packed blocks. A pass also starts early when a write queues fold work or a traversal meets a vertex with many unfolded keys. See also

.

sql

false

true, false

Whether to show the welcome message when starting the REPL.

start

none

A file path

Path to a SurrealQL file that will be imported when starting the server.

start

30s

A duration;

disables it

The interval at which a node looks for index builds that a node which stopped left unfinished, and resumes them.

start

5s

A duration

The interval at which to compact queued index updates. Writes to a full-text, count, or vector index enqueue pending updates that a background task folds into the index; this controls how often that task runs. One node in a cluster performs the work for the whole cluster. Lengthening the interval allows the pending backlog to grow, which increases the work each subsequent query must do to read through it.

start

none

A string 16, 24, or 32 bytes long

Encryption key to use for on-disk encryption. Not currently in use.

start

none

String to a path

Path to the CA file used when connecting to the remote KV store.

start

none

String to a path

Path to the certificate file used when connecting to the remote KV store.

start

none

String to a path

Path to the private key file used when connecting to the remote KV store.

start

false

true, false

Whether to lazy-load Surrealism modules instead of eagerly compiling them at server startup.

start, fix

info

none, full, error, warn, info, debug, trace

The logging level for the database server. In a release build,

and

give the same output as

, because the TRACE level is compiled only into debug builds.

start

false

true, false

Toggles file output.

start

text

text, json

The format for log file output.

start

none

none, full, error, warn, info, debug, trace

Override the logging level for file output. In a release build,

and

give the same output as

, because the TRACE level is compiled only into debug builds.

start

surrealdb.log

String to a file

Filename for logs (default: surrealdb.log)

start

logs

String to a path

Sets the directory for logs

start

daily

daily, hourly, never

Sets the rotation duration for logs.

start

text

text, json

Sets the format for logs.

start

SURREAL_LOG logging level

none, full, error, warn, info, debug, trace

Override the logging level for OpenTelemetry. In a release build,

and

give the same output as

, because the TRACE level is compiled only into debug builds.

start

none

String to a host:port

Send logs to the specified host:port

start

text

text, json

Set the format of the logs to the socket.

start

SURREAL_LOG logging level

none, full, error, warn, info, debug, trace

Override the logging level for socket logs. In a release build,

and

give the same output as

, because the TRACE level is compiled only into debug builds.

sql

main

A string

The namespace to connect to via the REPL.

export, import

none

A string

The namespace selected for the import/export operation.

start

false

true, false

Whether to hide the startup banner.

false

true, false

Whether to disable default namespace and database creation. Conflicts with SURREAL_DEFAULT_DATABASE and SURREAL_DEFAULT_NAMESPACE, which set a default value for namespace and database for a new instance.

start

false

true, false

Whether to suppress the server name and version headers.

start

15s

A duration

The interval at which to process and archive inactive nodes.

start

300s

A duration

The interval at which to process and cleanup archived nodes.

start

3s

A duration

The interval at which to refresh node registration information. The

endpoint reports the node unhealthy when its cluster heartbeat is older than three times this interval (9s by default). The multiple is fixed and cannot be configured. From SurrealDB 4.0, on a datastore that several nodes write, such as TiKV, a value above 10s stops the node from starting.

export, import, sql, start

none

A string

Database authentication password to use when connecting. For

, the password of the initial root user, which cannot be empty or only whitespace.

SURREAL_PATH

fix, mcp, start

memory

A string

Database path used for storing data. As a required argument (albeit with a default), it is not passed in via --path.

start

None (before 4.0.0, best-effort)

best-effort|compute-only|all-read-only

Deprecated, and has no effect from 4.0.0 as queries run on a single execution engine. The server still accepts the three values so that an existing configuration keeps starting, logs a deprecation warning and ignores them, and refuses to start on any other value. Remove it from your configuration. Before 4.0.0 it chose how far to use the query planner introduced in 3.0: the default used the new planner for read-only statements and fell back to the earlier compute planner on paths it did not implement, and compute-only skipped the new planner entirely.

start

none

A duration

The maximum duration that a batch of statements can run for. From 3.3.0 it is also a hard wall-clock limit on each HTTP

,

and

request and on each RPC method call. An HTTP

or

request that exceeds it is answered with

, and a

request with a GraphQL error. The

,

and

RPC methods are exempt, but the calls made inside the transaction they control are still limited individually.

start

60s

A duration

How often the background reaper scans for tombstoned namespace, database, and index data to physically delete after a

statement. On the 2.x line it is available from 2.7.0 and covers namespaces and databases only.

start

10m

A duration

Minimum age a removed namespace, database, or index must reach before its data is reclaimed. The effective grace is the maximum of this value and

on TiKV backends. Not available on 2.x.

start

none

A duration

A duration specifying the minimum execution time after which a log is made to indicate a slow query

start

none

Comma-separated strings

A comma-separated list of parameter names whose values are listed in slow query logs. No parameter value is logged unless its name is listed here. See

.

start

none

Comma-separated strings

A comma-separated list of parameter names whose values are left out of slow query logs, even when

also names them.

start

false

true, false

Whether strict mode is enabled on this database instance. Ignored since SurrealDB 3.0 after which strictness is defined per database instead of instance.

start

60s

A duration

How long each built-in startup datastore operation may keep retrying before the server gives up. It is a per-operation budget rather than a total for startup, and it covers initialising the default namespace and database, initialising the root credentials, and registering this node in the cluster. Raise it where the storage backend is slow to accept the first writes.

start

A

folder in the system temporary directory

String to a directory

Sets the directory for storing temporary database files, including the files that queries spill to when they pass their memory budget. Before 4.0.0 there was no default and the directory was used only by queries with

. A value that is not a directory is refused at startup. See

.

start

10m

A duration

The interval at which the TiKV MVCC garbage collector runs. Only used with a TiKV datastore.

start

10m

A duration

How far behind the current timestamp the TiKV garbage collection safepoint may sit. Only used with a TiKV datastore. The effective

is never shorter than this value.

start

60s

A duration

The interval at which stale TiKV transaction locks are resolved. Only used with a TiKV datastore.

export, import, sql

none

A string

Authentication token in JWT format to use when connecting.

start

none

A duration

The maximum duration that any single transaction can run for.

start

false

true, false

Whether to allow unauthenticated access.

export, import, sql, start

none

A string

Database authentication username to use when connecting. For

, the username of the initial root user, which cannot be empty or only whitespace.

start

none

String to a path

Path to the certificate file for encrypted client connections.

start

none

String to a path

Path to the private key file for encrypted client connections.

false

true, false

Whether to rate limit authentication attempts on the HTTP

and

endpoints per client address. Only

and

are accepted; any other value is ignored and the default applies.

60

A u32

How many authentication attempts a single client address may make before being throttled. Set to

to disable the limiter.

60

A u32

The sustained rate, in attempts per minute, at which a client address's budget refills. The default of 60 is one attempt per second. Set to

to disable the limiter.

16384

A usize

How many client addresses the limiter tracks at once. Unauthenticated callers write to this store, so it is a bounded cache: at capacity a new address evicts an existing entry rather than the store growing or the limiter switching off. An evicted address starts again from a full burst, so the sustained rate binds an address only while it stays tracked. Size this above the number of distinct addresses the instance authenticates. Set to

to disable the limiter.

Dynamically calculated via greater of ((system memory / 2) - 1 GiB) and 16MiB

A usize

The capacity of the block cache in bytes. The cache holds recently read data blocks in memory. From 4.0 a compaction does not add the blocks it reads to the cache, and it removes the cached blocks of the tables it merged, so the cache holds blocks of live tables only.

65_536 (64 KiB)

A usize

The block size in bytes.

true

true, false

Whether to enable value log separation.

4096

A usize

The maximum number of transactions in a single grouped commit batch. This prevents unbounded memory growth while still allowing large batches for efficiency. Larger batches improve throughput but increase memory usage and commit latency. From 4.0 the setting has no effect, because SurrealKV groups the syncs of concurrent commits itself, and the server logs a warning and ignores it.

5ms

A duration in nanoseconds

The maximum wait time in nanoseconds before forcing a grouped commit. Ensures that transactions do not wait indefinitely under low concurrency and balances commit latency against write throughput. From 4.0 the setting has no effect, because SurrealKV groups the syncs of concurrent commits itself, and the server logs a warning and ignores it.

12

A usize

Threshold for deciding whether to wait for more transactions. If the current batch size is greater or equal to this threshold (and below SURREAL_SURREALKV_GROUPED_COMMIT_MAX_BATCH_SIZE), then the coordinator will wait up to SURREAL_SURREALKV_GROUPED_COMMIT_TIMEOUT to collect more transactions. Smaller batches are flushed immediately to preserve low latency. From 4.0 the setting has no effect, because SurrealKV groups the syncs of concurrent commits itself, and the server logs a warning and ignores it.

Depends on system memory: 64 MiB below 1 GiB, 128 MiB below 4 GiB, 256 MiB below 16 GiB, 1 GiB below 64 GiB, and 4 GiB otherwise

A usize

The most memory a memtable holds before it is flushed to disk. The memtable holds recent writes. The limit counts the values the memtable holds as well as the memory area that indexes them, so a large load flushes to disk while it runs. From 4.0 it bounds the memory the memtable holds, not the size of a transaction: a commit too large for the memtable is written directly to a level-0 table. Before 4.0 a transaction larger than this fails, so raise it for large transactions.

false

true, false

Whether to enable versioned index. Only applies when versioning is enabled. From 4.0 the setting has no effect, because SurrealKV 1.0 reads versions without a separate index, and the server logs a warning and ignores it.

64 MiB (systems under 4 GiB), 128 MiB (up to 16 GiB), 256 MiB (up to 64 GiB), 512 MiB (others)

A usize

The maximum size of a value log file in bytes.

4096 (4 KiB)

A usize

The value log threshold in bytes. Values larger than this are stored in the value log.

1

A u8

Which TiKV cluster API version to use.

true

true, false

Whether to use asynchronous transactions.

none

A string

A string specifying the keyspace identifier for data isolation.

1,000,000

A u32

The most keys that one range delete in a TiKV transaction may remove. A larger range fails with an error rather than growing the transaction without bound.

true

true, false

Whether the background TiKV garbage collection runs. When

, the stale lock cleanup is skipped as well. The interval and lifetime are set with

and

, described under

.

4,194,304 (4 MiB)

A usize

Sets the maximum size of a gRPC message that SurrealDB receives from TiKV, which limits how much one read can return. A single record or a batch of records larger than this fails with

. Raise it together with

, so that a record that can be written can also be read back.

4,194,304 (4 MiB)

A usize

Sets the maximum size of a gRPC message that SurrealDB sends to TiKV, which limits the size of a single write. A record larger than this fails with an

. TiKV refuses a single write above its own entry limit of 8 MiB whatever this is set to, with

.

true

true, false

Whether the server checks that the TiKV cluster answers when it connects at startup, so that an unreachable cluster fails the start rather than the first query.

true

true, false

Whether to use one-phase transaction commit.

10

A usize

The duration in seconds for requests before they time out.

30

Seconds (integer)

How long a shutdown waits for running transactions to finish before it continues.

10

Seconds (integer)

The longest that the last garbage collection pass at shutdown may take, so that a slow pass does not hold up the exit.

none

A file path

Path to the CA certificate used to verify the TiKV and PD servers. Setting all three

variables switches the connection to mutual TLS (mTLS).

none

A file path

Path to the client certificate that SurrealDB presents during mTLS. It must be set together with the CA and key paths.

none

A file path

Path to the private key that matches the client certificate. It must be set together with the CA and certificate paths. If only some of the three paths are set, SurrealDB refuses to connect and reports which are missing.

true

false, true

Whether to enable background WAL file flushing. 2.x only: from 3.0.0, set a duration on

instead.

200 (milliseconds)

A usize

The interval in milliseconds between background flushes. 2.x only: from 3.0.0, set a duration on

instead.

0

A usize

Readahead size for blob compaction/GC.

snappy

none, snappy, lz4, zstd

Compression type used for blob files.

268,435,456 (256 MiB)

A usize

The target blob file size for RocksDB.

0.5

Float between 0 and 1

Fractional age cutoff for blob GC.

0.5

Float between 0 and 1

Discardable ratio threshold to force GC.

Dynamically calculated via greater of ((system memory / 2) - 1 GiB) and 16MiB

A usize

RocksDB

in bytes

65,536 (64 KiB)

A usize

The size of each uncompressed data block in bytes.

4 MiB (systems under 4 GiB), 8 MiB (up to 16 GiB), 16 MiB (others)

A usize

The readahead buffer size used during compaction.

level

level, universal

Use to specify the database compaction style.

false

true, false

Whether a graceful shutdown runs a full compaction of the database after flushing the memtables, as

does. On a large database this can take minutes or hours, which can be longer than the shutdown grace period of a container orchestrator.

50

A usize

The number of deletions to track in the window.

0.5

A float

The ratio of deletions to track in the window.

1000

A usize

The size of the window used to track deletions.

true

true, false

Whether to enable separate key and value file storage.

true

true, false

Whether to enable blob garbage collection for RocksDB.

true

true, false

Whether to use separate queues for WAL writes and memtable writes.

4

A usize

The number of files needed to trigger level 0 compaction.

2

A u64

The number of sequential reads an iterator makes on one SST file before RocksDB starts to prefetch. Set to

to prefetch from the first read.

4096

A usize

The maximum number of transactions in a single grouped commit batch. Used to prevent unbounded memory growth while still allowing large batches for efficiency. Larger batches improve throughput but increase memory usage and commit latency.

5ms

A duration

The maximum wait time in nanosecond before forcing a grouped commit. Used to ensure that transactions do not wait indefinitely when concurrency is low, and to balance between transaction latency and write throughput.

12

A usize

Threshold for deciding whether to wait for more transactions. If the current batch size is greater or equal to this threshold (and below ROCKSDB_GROUPED_COMMIT_MAX_BATCH_SIZE), then the coordinator will wait up to ROCKSDB_GROUPED_COMMIT_TIMEOUT to collect more transactions. Smaller batches are flushed immediately to preserve low latency.

8,192 (8 KiB)

A usize

The size of the first prefetch. Each following read doubles it, up to

.

Number of CPUs * 2

A usize

The maximum number of threads to use for flushing and compaction.

10

A usize

The maximum number of information log files to keep.

8

An i32

The number of level-0 files at which RocksDB starts to slow down writes. The RocksDB default is 20.

12

An i32

The number of level-0 files at which RocksDB stops writes until compaction reduces the count. Keep it above

. The RocksDB default is 36.

262,144 (256 KiB)

A usize

The largest prefetch an iterator grows to. A larger value can make long range scans faster, but uses more memory for each iterator, so many concurrent scans multiply the cost.

4

A usize

The maximum number threads which will perform compactions.

1024

A usize

The maximum number of open files which can be opened by RocksDB.

2 (systems under 4 GiB), 4 (up to 16 GiB), 8 (up to 64 GiB), 32 (others)

A usize

The maximum number of write buffers which can be used.

0.1

An f64

The size of the prefix bloom filter in each memtable, as a ratio of the write buffer size. RocksDB caps it at 0.25. Set to

to disable it. Only used when

is true.

4096

A usize

The minimum size in bytes of a value for it to be stored in blob files.

2

A usize

The minimum number of write buffers to merge before writing to disk.

3600

A u64

The interval in seconds after which RocksDB rewrites an SST file that no other compaction has touched. This removes old versions and tombstones from data that is rarely written. Set to

to disable.

true

true, false

Whether SST bloom filters are built on table and key-type prefixes, so that a scan can skip SST files that hold nothing from the table it reads.

2

A usize

The number of runtime worker threads kept free for asynchronous work while other transactions wait on a storage call. A larger value keeps more threads free under load, but lowers the number of storage calls that can run at once. See

.

true

true, false

Whether scans and counts verify the checksum of each block when it is first read from disk. Disabling it makes cold scans faster, and is only safe on trusted storage.

30

A u64

How long a graceful shutdown waits in seconds for a compaction that started after the final flush, so that the next start does not have to repeat it. Set to

to wait without a limit.

0

A usize

The maximum allowed space usage for SST files in bytes. The default of 0 means unlimited and disables space monitoring. When this limit is reached, the datastore enters read-and-deletion-only mode, where only read and delete operations are allowed. This allows gradual space recovery through data deletion.

warn

none, full, error, warn, info, debug, trace

The information log level of the RocksDB library.

67,108,864 (64 MiB)

• •

The target file size for compaction in bytes.

2

A usize

The target file size multiplier for each compaction level.

Number of CPUs on machine

A usize

The number of threads to start for flushing and compaction.

-1

An i32

The percentage of output files to compress.

compresses all of them. Only used when

is

.

4,294,967,295 (no limit)

A u32

The largest number of files in a single merge. Only used when

is

.

200

A u32

The largest acceptable storage amplification, as a percentage of the size of the live data. Only used when

is

.

2

A u32

The smallest number of files in a single merge. Only used when

is

.

1

An i32

How much larger, as a percentage, a file can be than the files before it and still be picked for the same merge. Only used when

is

.

total

similar_size, total

How RocksDB decides that a merge has collected enough files. Only used when

is

.

0

A usize

The write-ahead-log size limit in MiB.

true

true, false

Whether SST bloom filters also hold whole keys as well as prefixes. Whole keys make point lookups faster, at the cost of larger filters. Disabling it can suit a workload that mostly scans. Only used when

is true.

32 MiB (systems under 1 GiB), 64 MiB (up to 16 GiB), 128 MiB (others)

A usize

The amount of data each write buffer can build up in memory.

true

true, false

Whether to sync writes to disk before acknowledgement. 2.x only: replaced in 3.0.0 by

.

500

The maximum delay between transaction retries in milliseconds.

5

The maximum number of times a transaction can be retried.

5000

The maximum transaction timeout in milliseconds.