The official SurrealDB SDK for Python.

       

     

surrealdb.py

The official SurrealDB SDK for Python.

Documentation

View the SDK documentation here.

How to install

# Using pip
pip install surrealdb

# Using uv
uv add surrealdb

Quick start

In this short guide, you will learn how to install, import, and initialize the SDK, as well as perform the basic data manipulation queries.

This guide uses the Surreal class, but this example would also work with AsyncSurreal class, with the addition of await in front of the class methods.

Running SurrealDB

You can run SurrealDB locally or start with a free SurrealDB cloud account.

For local, two options:

1. Install SurrealDB and run SurrealDB. Run in-memory with:

surreal start -u root -p root

2. Run with Docker.

docker run --rm --pull always -p 8000:8000 surrealdb/surrealdb:latest start

Learn the basics

# Import the Surreal class
from surrealdb import Surreal, RecordID, Table

# Using a context manger to automatically connect and disconnect
with Surreal("ws://localhost:8000/rpc") as db:
    db.signin({"username": 'root', "password": 'root'})
    db.use("namepace_test", "database_test")

    # Create a record in the person table
    db.create(
        "person",
        {
            "user": "me",
            "password": "safe",
            "marketing": True,
            "tags": ["python", "documentation"],
        },
    )

    # Read all the records in the table
    print(db.select("person").execute())

    # Update all records in the table
    print(db.update("person", {
        "user":"you",
        "password":"very_safe",
        "marketing": False,
        "tags": ["Awesome"]
    }))

    # Delete all records in the table
    print(db.delete("person"))

    # You can also use the query method 
    # doing all of the above and more in SurrealQl
    
    # In SurrealQL you can do a direct insert 
    # and the table will be created if it doesn't exist
    
    # Create (sync query() returns a builder - call .execute() to run it)
    db.query("""
    insert into person {
        user: 'me',
        password: 'very_safe',
        tags: ['python', 'documentation']
    };
    """).execute()

    # Read - .first() returns the first statement's result (the rows)
    print(db.query("select * from person").first())
    
    # Update
    print(db.query("""
    update person content {
        user: 'you',
        password: 'more_safe',
        tags: ['awesome']
    };
    """).execute())

    # Delete
    print(db.query("delete person").execute())

CRUD builder pattern (v3.0)

create, update, upsert, delete, and insert return an awaitable (or lazy, for sync) builder. The builder exposes chainable clause methods that map directly to SurrealQL clauses.

from surrealdb import AsyncSurreal, RecordID, Table

async with AsyncSurreal("ws://localhost:8000/rpc") as db:
    await db.signin({"username": "root", "password": "root"})
    await db.use("ns", "db")

    # Sugar: db.create(record, data) is equivalent to .content(data)
    await db.create(RecordID("person", "tobie"), {"name": "Tobie"})

    # Or use the builder explicitly
    await db.create(RecordID("person", "tobie")).content({"name": "Tobie"})
    await db.update(RecordID("person", "tobie")).replace({"name": "Tobie"})
    await db.update(RecordID("person", "tobie")).merge({"vip": True})
    await db.update(RecordID("person", "tobie")).patch([
        {"op": "replace", "path": "/vip", "value": False},
    ])

    # `insert` accepts a `relation=True` kwarg or a chained `.relation()`
    await db.insert(Table("likes"), {"in": ..., "out": ...}, relation=True)
    await db.insert(Table("likes")).relation().content({"in": ..., "out": ...})

The builder is typed via @overload:

• RecordID target -> dict[str, Value]
• Table target -> list[Value]
• str target -> Value (a record-id string returns a dict; a table-name string returns a list - the type checker can't tell them apart, so falls back to Value)

select() returns a builder, like every other CRUD method - await it on the async client, .execute() it on the blocking one - and it unwraps single records:

• select(RecordID(...)) (or a "table:id" string) -> dict[str, Value] | None (None when the record does not exist)
• select(Table(...)) (or a bare table-name string) -> list[Value]

row = await db.select(RecordID("person", "tobie"))  # dict | None
rows = await db.select(Table("person"))             # list

What a RecordID and a Table accept

Both constructors check their arguments, so a mistake raises TypeError at the line that made it rather than coming back from the server as Parse error.

A table name must be a str, and that is the only rule — SurrealDB accepts any string, including one that is empty, has spaces, is unicode, starts with a digit, or contains a colon.

A record id must be one of str, int, uuid.UUID, list, tuple, dict, or a Range (see below). The union is exported as RecordIdValue if you want to annotate against it. Notably rejected: None, bool (Python's bool is an int, but SurrealDB has no boolean id), float, and bytes — the server refuses all of them.

RecordID("person", "tobie")     # ok
RecordID("person", ["a", 1])    # ok — composite ids are arrays or objects
RecordID(1, "tobie")            # TypeError: the arguments look swapped
Table(None)                     # TypeError: name must be a str

The check applies to values you construct. Records decoded from a response bypass it, so that a future server sending an id type this SDK does not know about still reads back — RecordID.id is therefore typed more widely than RecordIdValue, and is not guaranteed to be one of the types above.

Before 3.0.0, Table(None) was not rejected anywhere and db.insert(Table(None), rows) succeeded, writing to a table literally named None. If you have code that builds a table name dynamically, that is the case to check.

Selecting only some fields

Pass fields= to narrow the projection, so the server sends only what you ask for rather than the whole record:

await db.select(RecordID("person", "tobie"), fields=["name", "email"])
await db.select(Table("person"), fields=["address.city"])

A dot walks into a nested object. Each segment is escaped separately, so a field name containing a space or unicode is quoted correctly and a field list can never smuggle SurrealQL into the statement. A field whose name genuinely contains a dot cannot be spelled this way — use query() for that.

id is not included unless you ask for it, exactly as in SurrealQL, so a model passed to into= that declares an id field needs fields=["id", ...].

Anything beyond a list of field names — aliases, functions, WHERE, ORDER BY — is a query(), not a select().

Record ranges

A RecordID whose id is a Range targets every record in that range, and every CRUD method returns all of them — the same as the equivalent "person:1..=3" string:

from surrealdb import RecordID, Range, BoundIncluded, BoundExcluded

first_three = RecordID("person", Range(BoundIncluded(1), BoundIncluded(3)))
await db.select(first_three)            # every record in person:1..=3
await db.delete(first_three)            # deletes them all, returns them all
await db.select("person:1..=3")         # the same target, spelled as a string

Use BoundExcluded for .. rather than ..=, and None for an open end (Range(BoundIncluded(1), None) is person:1..).

Two caveats. A range needs a table, so a bare Range is not a resource target — db.select(Range(...)) raises SurrealError when the builder runs; wrap it in a RecordID. And the @overloads above resolve on the static type RecordID, which says nothing about the id, so a type checker still reads select(first_three) as dict | None while it returns a list at runtime. Cast, or use the string form, if you need the narrower type.

Mapping rows to a model (into=)

Pass the keyword-only into= argument to map each returned record onto a model class - a dataclass, a pydantic BaseModel, or any class whose constructor accepts the record's fields as keyword arguments. The return type is narrowed precisely per overload: a single-record target resolves to Model (or Model | None), a table target to list[Model].

from dataclasses import dataclass

@dataclass
class Person:
    id: RecordID
    name: str

# select: single record -> Person | None, table -> list[Person]
person = await db.select(RecordID("person", "tobie"), into=Person)  # Person | None
people = await db.select(Table("person"), into=Person)              # list[Person]

# create / update / upsert / delete map the written record(s) too
created = await db.create(RecordID("person", "tobie"), {"name": "Tobie"}, into=Person)
updated = await db.update(Table("person"), {"name": "Updated"}, into=Person)  # list[Person]

# insert maps the inserted records
inserted = await db.insert(Table("person"), [{"name": "A"}], into=Person)  # list[Person]

# the no-data builder form carries the model through its clause methods
p = await db.create(RecordID("person", "jaime"), into=Person).merge({"name": "Jaime"})

# map each ROW of a single query statement with into(Model, rows=True)
rows = await db.query("SELECT * FROM person").into(Person, rows=True)  # list[Person]

Sync connections take the same into= argument and run eagerly:

person = db.select(RecordID("person", "tobie"), into=Person).execute()  # Person | None
created = db.create(RecordID("person", "tobie"), {"name": "Tobie"}, into=Person)
rows = db.query("SELECT * FROM person").into(Person, rows=True)  # list[Person]

Omitting into= leaves the raw dict / list[Value] results completely unchanged.

The model has to accept every field of the records it is given. update(record, data) writes data as the record's whole content, so the example above leaves each person with just id and name — write a field Person does not declare and the mapping fails with an UnexpectedResponseError naming both sides:

into=Person could not be built from this record: Person.__init__() got an
unexpected keyword argument 'active'. The record has ['active', 'id']; Person
accepts ['id', 'name'].

Give the extra fields defaults, add them to the model, or SELECT only the columns it declares.

Sync usage has no await to defer to, so the rule is: a write you give its data runs immediately; everything else returns a builder. create/update/upsert/insert(target, data) run at once and return the result. select(), delete(), query() and the no-data write forms return a builder, which you run with .execute() (or a clause such as .merge(...)), or iterate to stream the rows. A builder never auto-executes on bool(), ==, indexing or attribute access, and len(...) or indexing one raises a TypeError that tells you to call .execute().

A builder you never run does nothing, so one that is thrown away unrun - a db.delete(Table("person")) missing its .execute(), or an async call missing its await - emits a RuntimeWarning naming the call and the line that made it, the way asyncio does for a coroutine that was never awaited.

from surrealdb import Surreal

with Surreal("ws://localhost:8000/rpc") as db:
    db.signin({"username": "root", "password": "root"})
    db.use("ns", "db")

    # Passing data runs immediately and returns the created record dict.
    tobie = db.create(RecordID("person", "tobie"), {"name": "Tobie"})

    # No-data form returns a builder; a terminal clause method runs it.
    out = db.create(RecordID("person", "alice")).merge({"name": "Alice"})

    # Clause-less run: call .execute() explicitly.
    empty = db.create(RecordID("person", "bob")).execute()

    # select() and delete() return a builder; .execute() runs it.
    row = db.select(RecordID("person", "tobie")).execute()  # dict | None
    db.delete(RecordID("person", "bob")).execute()

    # Iterating a builder streams its rows instead of loading them all.
    for person in db.select(Table("person")):
        print(person["name"])

    # query() returns a builder; call .execute()/.first()/.into().
    db.query("DELETE person;").execute()

On 3.x, DELETE names a table that has to exist: db.query("DELETE temp_data;") raises NotFoundError: The table 'temp_data' does not exist rather than deleting nothing. (2.x returns an empty result instead — see Talking to a SurrealDB 2.x server.) Use REMOVE TABLE IF EXISTS temp_data; when you cannot be sure.

Thread safety

The no-data sync builder guards its cache with a per-builder lock so calling .execute() from multiple threads issues exactly one RPC. It is not safe for concurrent reconfiguration though — calling .merge() on one thread while another calls .execute() is a race on the builder's clause/data state that the lock does not cover. Treat builders as single-shot, single-owner values; pass the realised result between threads, not the builder itself.

The underlying BlockingWsSurrealConnection is itself thread-safe (it serialises send/recv with an internal lock), so sharing a connection across threads and issuing per-thread operations against it is fine.

Async cancellation and server truth

If you cancel() an async task that's awaiting an in-flight builder, the SDK does the right thing on the client side: the cache is reset so fresh callers retry, and concurrent peer awaiters see a SurrealError rather than a phantom CancelledError they didn't request.

What it cannot do is roll back the server. Once an RPC has reached SurrealDB, the operation may still complete server-side even after the client cancels. For mutations this means cancellation is not an abort — re-read the affected records before assuming "nothing happened", or wrap mutations in a BEGIN ... COMMIT block via query() if you need atomic rollback semantics.

Multi-statement queries and transactions (issue #232 fix)

query() always returns a list[Value] - one entry per statement, even for a single statement - so multi-statement queries and BEGIN ... COMMIT blocks never silently drop results. Use .first() for the first statement's result (or None when there are no statements).

rows = await db.query("SELECT * FROM person")       # [people_list]
first = await db.query("SELECT * FROM person").first()  # people_list
many = await db.query(
    "SELECT * FROM person; SELECT count() FROM person GROUP ALL"
)
# many is [people_list, count_list]

# Sync: query() returns a builder - run it explicitly.
rows = db.query("SELECT * FROM person").execute()   # [people_list]
first = db.query("SELECT * FROM person").first()    # people_list

You can also map the N statement results onto a dataclass via .into():

from dataclasses import dataclass

@dataclass
class Stats:
    created: dict
    all_people: list
    count: int

result = await db.query(
    "CREATE person:tobie SET name = 'Tobie';"
    "SELECT * FROM person;"
    "SELECT count() FROM person GROUP ALL"
).into(Stats)

Or map each row of a single statement's result onto a model with .into(Model, rows=True), which returns list[Model]:

people = await db.query("SELECT * FROM person").into(Person, rows=True)  # list[Person]

For the raw server response (status, time, error per statement), keep using query_raw().

Client-side transactions and sessions

Multi-session and client-side transactions are supported only for WebSocket connections (ws:// or wss://). They are not available for HTTP or embedded connections.

async with AsyncSurreal("ws://localhost:8000/rpc") as db:
    await db.signin({"username": "root", "password": "root"})
    await db.use("ns", "db")

    # Create a session
    session = await db.new_session()
    await session.use("ns", "db")

    # Start a transaction on the session
    txn = await session.begin_transaction()
    await txn.create(RecordID("account", "alice"), {"balance": 100})
    await txn.update(RecordID("account", "bob")).merge({"balance": 50})

    # Commit (or call `await txn.cancel()` to roll back)
    await txn.commit()

    await session.close_session()

The same CRUD builder, query, and run() API is available on both AsyncSurrealSession / BlockingSurrealSession and AsyncSurrealTransaction / BlockingSurrealTransaction, along with query_raw(), info(), and version(). Every one of them scopes to the session (and transaction) it was called on, so you never pass session_id or txn_id by hand.

run() - calling SurrealDB functions

result = await db.run("fn::increment", [1])
greeting = await db.run("fn::greet", ["world"])

Live queries

Live queries let you subscribe to changes on a table and receive a notification whenever a record is created, updated, or deleted. They are a WebSocket-only feature (ws:// or wss://).

The API is three methods:

• live(table, diff=False) - start a live query on a table and return its UUID. Pass diff=True to receive JSON Patch diffs instead of full records.
• subscribe_live(query_uuid) - return a generator (async generator for the async client) that yields notification dicts. Each notification has an "action" ("CREATE", "UPDATE", or "DELETE") and a "result" (the affected record).
• kill(query_uuid) - stop a running live query.

You can also start a live query through query("LIVE SELECT * FROM ..."), which returns the same UUID you can pass to subscribe_live().

Async

import asyncio
from surrealdb import AsyncSurreal

async def main():
    # Connection that owns the subscription.
    async with AsyncSurreal("ws://localhost:8000/rpc") as db:
        await db.signin({"username": "root", "password": "root"})
        await db.use("ns", "db")

        live_id = await db.live("person")           # -> UUID
        subscription = await db.subscribe_live(live_id)

        # Drive the mutation on a SEPARATE connection (see caveats below).
        async with AsyncSurreal("ws://localhost:8000/rpc") as writer:
            await writer.signin({"username": "root", "password": "root"})
            await writer.use("ns", "db")
            await writer.create("person", {"name": "Jaime"})

        # Wait for the notification (guard with a timeout in real code).
        notification = await asyncio.wait_for(subscription.__anext__(), timeout=10)
        print(notification["action"])   # "CREATE"
        print(notification["result"])   # the created record

        await db.kill(live_id)

asyncio.run(main())

Blocking

from surrealdb import Surreal

with Surreal("ws://localhost:8000/rpc") as db:
    db.signin({"username": "root", "password": "root"})
    db.use("ns", "db")

    live_id = db.live("person")            # -> UUID
    subscription = db.subscribe_live(live_id)

    # Mutate on a SEPARATE connection so the notification can arrive.
    with Surreal("ws://localhost:8000/rpc") as writer:
        writer.signin({"username": "root", "password": "root"})
        writer.use("ns", "db")
        writer.create("person", {"name": "Jaime"})

    for notification in subscription:
        print(notification["action"], notification["result"])
        break                              # generator blocks for the next one

    db.kill(live_id)

Caveats

• Mutate on a separate connection. The connection that owns a subscription is busy receiving live notifications, so running CREATE/UPDATE/DELETE on that same connection races the query responses against the incoming notifications. Perform the mutations that should trigger notifications on a second connection (this is exactly what the test suite does).
• Blocking client: one subscriber per connection. The blocking subscribe_live() reads notifications straight off the socket, so a single blocking connection supports only one concurrent subscriber. Use a separate connection per live subscription (or the async client, which fans notifications out to per-subscriber queues).

Streaming queries

You get this for free. Against SurrealDB v3.3.0 or later over a WebSocket, query() already asks for its answer as a stream of frames and rebuilds it as it arrives - and so do select(), create(), upsert() and every builder, because they all go through the same call. The answer is identical; what changes is that the server no longer has to finish before any of it reaches you, and there is no single enormous response to decode.

Nothing to switch on, and nothing to change in your code:

people = await db.query("SELECT * FROM person")   # streamed, if the server can

.rows() on any builder is the visible half, for when you want the rows as they arrive rather than the whole answer at the end:

async for person in db.query("SELECT * FROM person").rows():
    ...

Iterating the builder itself is the same thing without the .rows(), on the async client as on the blocking one - await builder for the whole answer, async for row in builder for the rows as they arrive:

async for person in db.select(Table("person")):
    ...

It needs the same v3.3.0 server and a WebSocket connection. Against an older server, or over HTTP, the query runs the buffered way and its rows are handed back one at a time, so the call works everywhere - see Where it streams below.

Nothing is sent until you start iterating.

Two views over one answer

Iterate the stream for rows, as they arrive:

async for person in db.query("SELECT * FROM person").rows():
    print(person["name"])

or call .statements() for one completed result per statement, which is the shape query() returns:

async for statement in db.query(
    "SELECT * FROM person; SELECT count() FROM person GROUP ALL"
).statements():
    print(statement.index, statement.value)

Each StatementResult carries index, value, time (the server's own timing, verbatim), query_type ("live" for a LIVE SELECT, otherwise None), and single - true when value is one bare value rather than a list of rows, as for RETURN 1 + 2 or SELECT ... FROM ONLY.

A stream is read once, and both views draw from the same frames, so pick one per call.

Frames, when neither view will do

.stream() is the low-level view, and almost always the wrong one to start with. It yields every value, error and completion as the frame that says which, each carrying the index of the statement it belongs to:

from surrealdb import DoneFrame, ErrorFrame, ValueFrame

async for frame in db.query("SELECT * FROM person; THROW 'nope'; RETURN 42").stream():
    if isinstance(frame, ValueFrame):
        print(frame.index, frame.value)
    elif isinstance(frame, ErrorFrame):
        print(frame.index, "failed:", frame.error)
    elif isinstance(frame, DoneFrame):
        print(frame.index, "complete", frame.result.time)

Frames are for the one thing rows() and statements() cannot show: one statement failing while the statements after it carry on. Both of those stop at the first failure; here the THROW arrives as an ErrorFrame and RETURN 42 still produces its value and its DoneFrame.

The same rules apply as everywhere else. A value is provisional until its statement's DoneFrame, and an ErrorFrame retracts the values already yielded for that index - the difference is that you see it happen rather than having iteration raise. Statements are counted by their DoneFrames. A query that could not be completed at all, such as one whose connection died, still raises.

Rows as models

into= maps each row as it arrives, which is the case streaming is actually for - a large table read one model at a time, never held whole:

async for person in db.select("person").rows(into=Person):
    print(person.name)

The same on the blocking client, with with instead of async with:

for person in db.select("person").rows(into=Person):
    print(person.name)

The rows are typed, so a type checker sees what each loop yields. A record is dict[str, Value] and into=Model gives Model, however the target is spelled (Table, RecordID or a string), and the same goes for iterating the builder directly:

for row in db.select(Table("person")):                # dict[str, Value]
    ...
for person in db.select(Table("person"), into=Person):  # Person
    print(person.name)

A query's rows can be any value, so query(...).rows() yields Any unless you pass into=. What rows() returns is a SyncRows[T] (or AsyncRows[T]) - the iterable that also works as a context manager - which is the name to annotate with if you pass one around.

Rows are provisional until iteration ends

A row is delivered before the statement that produced it has finished - that is the point - so a statement that fails after emitting rows raises, and the rows it already yielded are void:

try:
    async for row in db.query("SELECT * FROM person; THROW 'nope'").rows():
        rows.append(row)          # these arrive, then the THROW raises
except SurrealError:
    rows.clear()                  # what arrived was never final

.statements() narrows this rather than removing it: it hands over a statement's value only once the server has called that statement final, so a statement you have received is settled - but a later statement can still fail, and earlier ones have already been yielded. For all-or-nothing across the whole query, use query(), or wrap the statements in BEGIN/COMMIT so the server rolls them back together.

Stopping early

Use async with (or with, or call aclose() / close()) so that stopping early tells the server to abandon the query instead of running it to completion:

async with db.query("SELECT * FROM huge_table").rows() as stream:
    async for row in stream:
        if found(row):
            break                 # the server stops here

Closing is not merely tidiness: an abandoned stream is still executing on the server, holding one of the connection's stream slots. Abandoning one without closing does still clean up once you let go of the iterator - the blocking client cancels at that moment, the async client on the event loop's next pass over finalisable generators. But letting go is the part that is easy to miss: a break inside a function that keeps running leaves the generator referenced by that frame, so the cancel waits for the frame to go, and the query it was meant to stop runs to completion in the meantime. Only closing stops the query at a moment you choose.

Blocking client

Identical, minus the as:

with db.query("SELECT * FROM person").rows() as stream:
    for person in stream:
        print(person["name"])

Frames only arrive while somebody reads the socket, so a blocking stream is advanced by the thread iterating it. It takes the connection lock in short slices rather than holding it, so other callers on the same connection keep working while a stream is open.

Where it streams

	Behaviour
WebSocket, server v3.3.0+	Streams. Rows arrive as they are produced.
WebSocket, older server	Runs the query the buffered way. Learned once per connection.
WebSocket, the query_stream RPC denied	Same, with its own reason - the operator denied streaming, not querying.
WebSocket, at the concurrency cap	query() is buffered for this query only, and it is not remembered: the cap is transient. An explicit .rows() raises instead.
Inside a client transaction	query() is never streamed - see the caveats below.
HTTP	Buffered - HTTP carries one response per request.
Embedded	Buffered.

For the streaming query() does on your behalf, every one of those is a retry rather than an error, and it is a refusal from the server that makes it safe: begin is framed before execution starts, so a refusal with no frame behind it means the query never ran, and asking again cannot run it twice. An explicit .rows() retries the first three rows the same way, but raises at the concurrency cap rather than quietly going buffered.

A socket that dies before the first frame is not such a refusal, and is not retried. begin may have been framed and the query may have been executing as the connection went, so nothing about it is proven - and retrying gains nothing anyway, because the server-side session dies with the socket. It raises the connection error instead, exactly as a buffered query on a dying socket always has. Retrying it used to report Specify a namespace to use, which blamed the query for a dead connection.

To take the invisible path off a whole connection:

db = AsyncSurreal("ws://localhost:8000/rpc", streaming=False)

That switches off the streaming query() does on your behalf. It does not override an explicit .rows() call, which streams whenever the server can - asking for a stream outright is taken as meaning it.

For query() the fallback costs nothing - the answer is buffered either way. For .rows() it gives up the two things streaming is for: rows do not arrive early, and the whole result is held in memory. When that matters, pass require_streaming=True and get an UnsupportedFeatureError instead of a quiet buffered answer:

async for row in db.query(sql).rows(require_streaming=True):
    ...

Feature detection is the protocol's own, and no version string is parsed: a server without query_stream answers Method not found, and one whose capabilities deny it answers Method not allowed. Both mean this connection will not stream, both are remembered after one request, and require_streaming=True reports which of the two it was - upgrading fixes the first, not the second.

Caveats

• Client-side buffering is bounded, not absent. The protocol has no per-stream flow control, so the only way to slow the server is to stop reading the socket - which the async client now does once a stream is far enough ahead, and undoes the moment anything else on the connection needs the reader. A slow consumer therefore holds a bounded number of frames rather than the rest of the answer, at the cost of the server pausing mid-query. The blocking client reads only on demand and never held anything to begin with.
• One task, or one thread, per stream. A stream is driven by whoever iterates it, and Python will not let two do so at once: closing an async generator while another task is awaiting a row from it raises RuntimeError: aclose(): asynchronous generator is already running, and CPython refuses to run one generator from two threads. To stop a stream another task is waiting on, cancel that task and await it before closing.
• Concurrent streams are capped. A connection allows 32 in flight by default (SURREAL_WEBSOCKET_MAX_CONCURRENT_STREAMS on the server); the 33rd is refused with a ValidationError saying so.
• LIVE SELECT in a stream. The statement's value is the live-query id, with query_type == "live". Notifications begin after the stream ends, and - as with live() - anything that happens before you call subscribe_live() is not delivered, so subscribe promptly.
• A failed statement stops the rest of the query. When a statement fails, iteration raises and the server is asked to abandon what is left - so statements after the failure may never run, where query() executes the whole query before raising. The difference only shows when the remainder is slow enough for the cancel to land, and it applies to side effects, not just results: CREATE a; THROW 'x'; SLEEP 3s; CREATE b leaves both records via query() and only a via .rows(). Wrap the statements in BEGIN/COMMIT if you need all-or-nothing.
• query() inside a client transaction is never streamed. Requests on one connection are served concurrently, so a commit could arrive while a streamed query was still executing and commit a prefix of it. Since nobody asked for a stream, the hazard is simply removed: anything from begin() takes the buffered path.
• .rows() inside a transaction still streams, because asking for a stream outright is taken as meaning it - but the hazard above is now yours to avoid. The stream runs on that transaction, so finish it before committing. The commit has to land while the server is still executing for this to bite, which a fast query usually finishes before; when it does bite it takes a prefix. Measured on UPDATE ... RETURN AFTER; SLEEP 3s; UPDATE ... with the commit sent during the sleep: the first UPDATE was committed, the second never ran, and the stream raised Couldn't update a finished transaction.

None, Null, and empty values

SurrealDB has two ways for a field to hold nothing, and they are different values:

SurrealQL	meaning	Python
NONE	the field is not there	None
NULL	the field is there and is null	surrealdb.Null

An unset option<T> column is NONE, and a NONE field does not appear in a record at all when you read it. NULL is a value the field holds, and a column has to be declared to permit it — option<T> alone does not, and rejects NULL.

from surrealdb import Null

db.create(rec, {"age": None})    # age is NONE — what option<int> expects
db.create(rec, {"age": Null})    # age is NULL — rejected by option<int>

This matters most when you read a record and write it back. A NULL field reads as Null, and sending Null back writes NULL again, so the round trip keeps the field:

row = db.select(rec).execute()   # {"nickname": Null}
row["name"] = "new name"
db.update(rec, row)              # nickname is still NULL

Null is falsy, like None, so if not row["nickname"] reads the way you would expect. It is deliberately not equal to None — the two are different values to the server, and treating them as one is what used to make the write-back above delete the field.

Before 3.0.0-beta.5 a NULL field decoded to None, which encoded back to NONE — so an ordinary read-modify-write silently removed every NULL field it touched. If you have code comparing a database value with is None, check whether the column can be NULL.

Sets read back as SurrealSet. SurrealDB's set<T> is a deduplicated sequence that accepts any member type, including set<object> and set<array> — which a Python set cannot hold, because its members must be hashable.

SurrealSet is a list subclass, so it indexes and compares like the sequence it is, and it encodes back under the set tag, so writing a record back keeps the field a set:

row = db.select(rec).execute()         # {"tags": SurrealSet(['a', 'b'])}
row["name"] = "new name"
db.update(rec, row)                    # tags is still a set

db.select(rec).execute()["tags"] == ["a", "b"]   # True — it is a list

Writing a plain Python set still works and is still sent as a set. The order is whatever the server sent: SurrealDB normalises a set, so <set>[3,1,2] comes back as [1, 2, 3].

Decoding a set to a plain list — as 3.0.0-beta.5 briefly did — loses the type on the way back: a schemafull set<T> field rejects the array outright, and a schemaless one silently becomes an array and stops deduplicating.

Sets need SurrealDB 3.x. 2.x has no CBOR set representation at all — see Talking to a SurrealDB 2.x server.

Error handling

Every error the SDK raises derives from SurrealError, so a single except SurrealError covers all of them regardless of transport.

Below that base, errors split into two branches that mean different things:

Branch	Meaning	Retry?
ServerError	The server ran your request and rejected it	No — it will fail again
TransportError	The request never produced a structured server response	Maybe — it may succeed

ServerError mirrors SurrealDB's structured error format, so you can match on kind and read typed details rather than parsing messages:

from surrealdb import Surreal
from surrealdb.errors import NotFoundError, ServerError, SurrealError, TransportError

with Surreal("ws://localhost:8000/rpc") as db:
    db.signin({"username": "root", "password": "root"})
    db.use("ns", "db")

    try:
        db.query("SELECT * FROM nonexistent:1").execute()
    except NotFoundError as error:
        print(error.kind, error.table_name)   # NotFound nonexistent
    except ServerError as error:
        print(error.kind, error.details)
    except TransportError as error:
        print("could not reach the server:", error)

Permissions do not raise

A signed-in record user whose PERMISSIONS deny a statement does not get an error. The server filters instead, the way row-level security does elsewhere: the statement comes back OK with an empty result, query_raw() reports status: "OK", and nothing is raised.

db.signup({"namespace": "ns", "database": "db", "access": "account", "variables": {...}})
db.use("ns", "db")

db.query("CREATE thing SET v = 1").execute()        # [[]] - denied, nothing was written
db.select(Table("thing")).execute()                 # []   - rows you cannot see are omitted
db.query("UPDATE thing:theirs SET v = 9").execute() # [[]] - denied, nothing was changed

This is deliberate, and the SDK does not turn it into an error. An error that only appeared when a record exists but is not yours would tell the caller the record exists, and one that depended on a WHERE in a permission rule would leak what that rule hides. It could not be told apart from an ordinary empty result in any case: a denied UPDATE of an existing record and an UPDATE of one that does not exist come back identically.

Two things follow when you are debugging:

• An empty result from a write is ambiguous. A write returns the rows it touched as the user could select them. If the user may CREATE a record but not SELECT it back, the record is created and the result is still empty. Check the outcome by reading the record with a connection that can see it, not by the write's own result.
• Nothing was written does not mean nothing was denied. An empty result tells you the statement produced no visible rows. To find out which rule is responsible, test the same statement as a user the rule lets through, or review the table's PERMISSIONS clause.

A root user is not subject to table permissions, so the same statement succeeds for them - which is also the quickest way to confirm a rule, and not the query, is what filtered the result.

Talking to a SurrealDB 2.x server

Six things behave differently against SurrealDB 2.x, all because of what the 2.x server does rather than anything the SDK chooses.

Error kinds need 3.x. The subclasses below ServerError come from the kind the server reports, and 2.x does not send one — its error payload has only a generic -32000 code and a message string. So every server-side failure arrives as InternalError:

	SurrealDB 3.x	SurrealDB 2.x
db.query("SELECT * FROM")	ValidationError	InternalError
db.query("THROW 'nope'")	ThrownError	InternalError

except SurrealError and except ServerError work on every version. A narrower except ThrownError matches only on 3.x and will silently not match on 2.x, so if you support both, catch the branch rather than the leaf, or read the message with str(error). The SDK cannot recover the classification — it is not sent.

Python sets need 3.x. Sets are encoded with SurrealDB 3.x's CBOR set tag. 2.x has no set representation at all — it returns SurrealQL sets as plain arrays — and rejects anything carrying the tag. Send a list instead when targeting 2.x; a list is what 2.x uses for a set<…> field anyway, and the server deduplicates it. On 3.x a list is not interchangeable with a set: a set<…> field rejects one.

let() wins over a query's own variables on a 2.x websocket. Passing a variable to query() shadows a let() binding of the same name for that one query — on 3.x, and on HTTP against any version, where the SDK replays let() bindings itself. On a 2.x websocket let() is a server-side session binding and that server resolves it the other way round:

db.let("limit", 99)
db.query("RETURN $limit", {"limit": 5})   # 5, except on a 2.x websocket: 99

The SDK sends the same request in both cases, so there is nothing for it to fix without knowing the server version up front. Use distinct names for session bindings and per-query variables if you target 2.x over a websocket.

A 2.x function body cannot read a variable from outside itself. On 3.x a stored function resolves $v from the session or from the query's parameters; on 2.x it resolves neither, on any transport:

db.query("DEFINE FUNCTION fn::readv() { RETURN $v; };").execute()
db.let("v", 42)
db.run("fn::readv")        # 42 on 3.x, None on 2.x

Pass the value as an argument (db.run("fn::add", [1, 2])) if you target 2.x — arguments work on every version.

DELETE on a table that does not exist is an error only on 3.x. 3.x raises NotFoundError: The table 'temp_data' does not exist; 2.x returns an empty result, as though it had deleted nothing:

db.query("DELETE temp_data;").execute()   # [[]] on 2.x, NotFoundError on 3.x

REMOVE TABLE IF EXISTS temp_data; behaves the same on both.

A 2.x server does not tell subscribers that a live query was killed. On 3.x the server sends a KILLED notification and subscribe_live() ends, whoever killed the query. On 2.x nothing is sent, so a subscriber to a query killed from another connection simply stops hearing anything and keeps waiting — there is no signal for the SDK to end the generator on. Calling kill() on the same connection you subscribed from ends the subscription on every version.

The TransportError branch is ConnectionUnavailableError (the host was unreachable or the socket closed), TransportTimeoutError (the request timed out), and HttpStatusError (a non-2xx HTTP response, carrying .status, .body, and .url). Each keeps the underlying library exception as __cause__ when you need the original detail.

Note that a non-2xx HTTP response is reported as HttpStatusError rather than a ServerError subclass, because SurrealDB answers those at the HTTP layer with a plain-text or JSON body and no structured RPC error to map. An invalid bearer token over HTTP, for example, surfaces as HttpStatusError with .status == 401, not NotAllowedError.

Invalid inputs are still rejected with plain ValueError / TypeError before anything is sent, following normal Python convention.

Migrating from 2.x

v3.0 is a breaking change. Highlights:

2.x	3.0
db.merge(record, data)	db.update(record).merge(data)
db.patch(record, data)	db.update(record).patch(data)
db.insert_relation(table, data)	db.insert(table, data, relation=True)
db.query("RETURN 1") -> single result	db.query("RETURN 1") -> [result] (use .first() / [0])
db.query("RETURN 1; RETURN 2") -> first result	db.query("RETURN 1; RETURN 2") -> list of all results
n/a	db.run("fn::name", [args])
n/a	db.query("...").into(MyDataclass)
Sync db.query("DELETE foo") runs immediately	Sync db.query("DELETE foo").execute() (returns list)
Sync db.create(rec)[...] (magic auto-exec)	Sync db.create(rec, data) eager, or db.create(rec).execute()
db.select(RecordID(...)) -> [record]	db.select(RecordID(...)) -> a builder; await/.execute() for the record dict or None
db.delete("my-table") (silently inlined)	db.delete(Table("my-table")) (raw string rejected)
A NULL field read as None	A NULL field reads as Null (None still means NONE)
set<T> read as a Python set	set<T> reads as a SurrealSet (writing a set is unchanged)

Bare-string resource targets are now strictly validated against the safe-identifier pattern ([A-Za-z_][A-Za-z0-9_]*) so user-supplied strings can never be concatenated into the generated SurrealQL. Names with hyphens, spaces, or other special characters must be wrapped in Table(...) or RecordID(...), both of which are parameter-bound.

Embedded Database

SurrealDB can also run embedded directly within your Python application natively. This provides a fully-featured database without needing a separate server process.

Installation

The embedded database ships as an optional native extension, surrealdb-embedded, so it is not part of the default install. Request it with the embedded extra:

pip install 'surrealdb[embedded]'

Or install using uv:

uv add 'surrealdb[embedded]'

Without the extra, an embedded URL raises UnsupportedEngineError telling you to install it - the remote http://, https://, ws:// and wss:// connections work either way.

Embedded connections do not authenticate: there is no server and no root user, so calling signin() on one raises NotAllowedError. Use use() to select a namespace and database and start querying.

For source builds, you'll need Rust toolchain and maturin:

uv run maturin develop --release --manifest-path embedded/Cargo.toml

In-Memory Database

Perfect for embedded applications, development, testing, caching, or temporary data.

import asyncio
from surrealdb import AsyncSurreal

async def main():
    # Create an in-memory database (can use "mem://" or "memory")
    async with AsyncSurreal("memory") as db:
        await db.use("test", "test")
        
        # Use like any other SurrealDB connection
        person = await db.create("person", {
            "name": "John Doe",
            "age": 30
        })
        print(person)
        
        people = await db.select("person")
        print(people)

asyncio.run(main())

File-Based Persistent Database

For persistent local storage:

import asyncio
from surrealdb import AsyncSurreal

async def main():
    async with AsyncSurreal("file://mydb") as db:
        await db.use("test", "test")
        
        # Data persists across connections
        await db.create("company", {
            "name": "Acme Corp",
            "employees": 100
        })
        
        companies = await db.select("company")
        print(companies)

asyncio.run(main())

Blocking (Sync) API

The embedded database also supports the blocking API:

from surrealdb import Surreal

# In-memory (can use "mem://" or "memory")
with Surreal("memory") as db:
    db.use("test", "test")
    
    person = db.create("person", {"name": "Jane"})
    print(person)

# File-based
with Surreal("file://mydb") as db:
    db.use("test", "test")
    
    company = db.create("company", {"name": "TechStart"})
    print(company)

When to Use Embedded vs Remote

Use Embedded (memory, mem://, file://, or surrealkv://) when:

• Building desktop applications
• Running tests (in-memory is very fast)
• Local development without server setup
• Embedded systems or edge computing
• Single-application data storage

Use Remote (ws:// or http://) when:

• Multiple applications share data
• Distributed systems
• Cloud deployments
• Need horizontal scaling
• Centralized data management

For more examples, see the examples/embedded/ directory.

Sessions in detail

• Sessions: Call attach() on a WS connection to create a new session (returns a UUID). Use new_session() to get an AsyncSurrealSession or BlockingSurrealSession that scopes all operations to that session. Call close_session() on the session (or detach(session_id) on the connection) to drop it.
• Transactions: On a session (or the default connection - though typical practice is to start on a session), call begin_transaction() to obtain a Transaction whose builder calls all participate in the same transaction. Call commit() to apply, or cancel() to roll back.

On HTTP or embedded connections, attach(), detach(), begin(), commit(), cancel(), and new_session() raise UnsupportedFeatureError with a message that sessions/transactions are only supported for WebSocket connections.

Observability with Logfire

Pydantic Logfire provides automatic instrumentation for SurrealDB operations, giving you instant observability into your database interactions. Logfire exports standard OpenTelemetry spans, making it compatible with any observability platform.

Quick start

Install Logfire using pip:

pip install logfire

Or install using uv:

uv add logfire

Usage:

import logfire
from surrealdb import AsyncSurreal

# Configure Logfire
logfire.configure()

# Instrument all SurrealDB operations
logfire.instrument_surrealdb()

# All database operations are now automatically traced
async with AsyncSurreal("ws://localhost:8000") as db:
    await db.signin({"username": "root", "password": "root"})
    await db.use("test", "test")
    
    # These operations will appear as spans in your traces
    await db.create("person", {"name": "Alice"})
    await db.query("SELECT * FROM person")

Features

• Automatic tracing: All database methods are instrumented automatically
• Smart parameter logging: Sensitive data (tokens, passwords) are automatically scrubbed
• OpenTelemetry compatible: Works with Jaeger, DataDog, Honeycomb, and other OTel platforms
• Minimal overhead: Efficient instrumentation with negligible performance impact
• Works with all connection types: HTTP, WebSocket, and embedded databases

Learn More

For a complete example with configuration options and best practices, see examples/logfire/.

WebSocket keepalive

Websocket connections send a keepalive ping on an idle socket and close the connection if the pong does not come back. Both intervals are in seconds, and both accept None to switch that half off:

from surrealdb import Surreal

# the defaults
db = Surreal("ws://localhost:8000/rpc", ping_interval=30, ping_timeout=10)

# behind a proxy with a short idle timeout
db = Surreal("ws://localhost:8000/rpc", ping_interval=10)

# let something else handle liveness
db = Surreal("ws://localhost:8000/rpc", ping_interval=None)

A peer that goes away is noticed within roughly ping_interval + ping_timeout, so the defaults detect it in about 40 seconds. SurrealDB also pings the client every 5 seconds of its own accord, which the library answers automatically — so the socket is rarely idle for long in either direction.

Both options are ignored by the HTTP and embedded transports, which have no websocket, exactly as streaming is. They require websockets 15.0 or newer, which is the SDK's declared floor.

Files

SurrealDB can store files in a bucket - in memory, on disk, or on object storage such as S3. A File is a reference to one: a bucket plus a key. It holds no bytes and is not a Python file object.

Buckets are currently an experimental server feature, so the server has to be started with SURREAL_CAPS_ALLOW_EXPERIMENTAL=files, and a bucket has to exist:

DEFINE BUCKET images BACKEND "memory";

from surrealdb import File, Surreal

with Surreal("ws://localhost:8000/rpc") as db:
    db.signin({"username": "root", "password": "root"})
    db.use("ns", "db")

    avatar = File("images", "/photos/avatar.png")

    with open("avatar.png", "rb") as handle:
        db.files.put(avatar, handle.read())  # bytes, not the handle itself

    png_bytes = db.files.get(avatar)         # -> bytes, or None if absent

    meta = db.files.head(avatar)             # -> FileMetadata | None
    print(meta.size, meta.updated)

    for entry in db.files.list("images", prefix="/photos", limit=50):
        print(entry.file.key, entry.size)

db.files is available on the four remote connection classes — websocket and HTTP, blocking and async — and on sessions and transactions opened from them. The async classes take the same calls with await. The two embedded classes inherit db.files, but the embedded engine does not enable the experimental files capability, so every call reports that.

File writes are not transactional. txn.files.put(...) is routed through the transaction's session, but file::* writes to the bucket backend rather than the transaction, so the write is visible to other connections immediately and survives txn.cancel() — while a CREATE in the same transaction rolls back as you would expect. That is SurrealDB's behaviour rather than the SDK's, and it is worth knowing before you pair a file write with a record write.

Method	Returns
put(file, content)	None, replacing anything at that key
put_if_not_exists(file, content)	None; writes only if the key is free
get(file)	bytes, or None if the file does not exist
exists(file)	bool
delete(file)	None; deleting an absent file is not an error
head(file)	FileMetadata, or None if absent
list(bucket, *, limit=, prefix=, start=)	list[FileMetadata]; start is exclusive
copy(source, target) / copy_if_not_exists	None
rename(file, key) / rename_if_not_exists	None; key is a str, within the same bucket

A key is normalised to start with /, so File("b", "a.txt") and File("b", "/a.txt") are the same file.

The two halves are different types: a File is the reference (bucket plus key), and the contents are plain bytes. put accepts bytes, bytearray or memoryview and get returns bytes. A str is refused rather than encoded for you — the server stores bytes, so accepting one would make put(f, s) succeed and get(f) == s be False. Encode it yourself, and pick the encoding:

db.files.put(readme, "hello".encode("utf-8"))
db.files.put(photo, handle.read())            # not the handle itself

Why files are bound, not written into queries

SurrealQL's file literal - f"bucket:/key" - accepts only [A-Za-z0-9_-./], and there is no escape mechanism. A key containing a space, an accent or an emoji cannot be written as a literal at all:

Parse error: Unexpected character ` `, file strings key's only allow alpha
numeric characters and `_`, `-`, `.`, and `/`

Every db.files method binds the File as a query parameter instead, which carries any key. str(File(...)) renders the literal form for logging, but is display-only - use File.is_literal_safe() if you need to know whether a given file could be written into a query, and bind it rather than interpolating it either way.

Agent memory

Agent memory is an optional client for Spectron, a memory service. It ships as its own distribution so it can move at its own pace, and is installed through an extra:

pip install 'surrealdb[memory]'

# Using uv
uv add 'surrealdb[memory]'

from surrealdb.memory import AsyncMemory, Memory

with Memory(
    context="acme-prod",
    endpoint="https://api.spectron.example",
    api_key="sk-spec-...",
) as memory:
    memory.remember("I work at Acme as CTO")
    hits = memory.recall("what do I do at Acme")
    print(hits.hits)

Memory is synchronous (backed by requests); AsyncMemory is the await-able equivalent (backed by aiohttp). Errors derive from MemoryServiceError — not MemoryError, which is a Python builtin. See memory/README.md for the full client documentation.

Without the extra, surrealdb.memory still imports, and any attribute names the missing package:

ImportError: 'Memory' needs the agent memory client, which ships separately.
    pip install 'surrealdb[memory]'    # or: uv add 'surrealdb[memory]'

Because it is a separate distribution, its version moves independently of the SDK's — you can hold surrealdb at 3.x and upgrade surrealdb-memory across its own majors.

Contributing

Contributions to this library are welcome! If you encounter issues, have feature requests, or want to make improvements, feel free to open issues or submit pull requests.

If you want to contribute to the Github repo please read the general contributing guidelines on concepts such as how to create a pull requests here.

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.