diff --git a/src/btrfs2s3/action.py b/src/btrfs2s3/action.py
index 6931445..c533dc8 100644
--- a/src/btrfs2s3/action.py
+++ b/src/btrfs2s3/action.py
@@ -14,6 +14,7 @@
 
 from btrfs2s3._internal.util import NULL_UUID
 from btrfs2s3._internal.util import SubvolumeFlags
+from btrfs2s3.stream_uploader import upload_non_seekable_stream_via_tempfile
 from btrfs2s3.thunk import Thunk
 from btrfs2s3.thunk import ThunkArg
 
@@ -26,6 +27,8 @@
 
 _LOG = logging.getLogger(__name__)
 
+DEFAULT_PART_SIZE = 5 * 2**30
+
 
 def create_snapshot(*, source: Path, path: Path) -> None:
     """Create a read-only snapshot of a subvolume.
@@ -94,12 +97,16 @@ def create_backup(  # noqa: PLR0913
     send_parent: Path | None,
     key: str,
     pipe_through: Sequence[Sequence[str]] = (),
+    part_size: int = DEFAULT_PART_SIZE,
 ) -> None:
     """Stores a btrfs archive in S3.
 
     This will spawn "btrfs -q send" as a subprocess, as there is currently no way
     to create a btrfs-send stream via pure python.
 
+    This will temporarily store the "btrfs send" stream on disk. A maximum of
+    part_size bytes will be stored at a time.
+
     Args:
         s3: An S3 client.
         bucket: The bucket in which to store the archive.
@@ -109,6 +116,9 @@ def create_backup(  # noqa: PLR0913
         key: The S3 object key.
         pipe_through: A sequence of shell commands through which the archive
             should be piped before uploading.
+        part_size: For multipart uploads, use this as the maximum part size.
+            Defaults to the well-known maximum part size for AWS, which is
+            currently 5 GiB.
     """
     _LOG.info(
         "creating backup of %s (%s)",
@@ -132,7 +142,13 @@ def create_backup(  # noqa: PLR0913
     # https://github.com/python/typeshed/issues/3831
     assert pipeline_stdout is not None  # noqa: S101
     try:
-        s3.upload_fileobj(pipeline_stdout, bucket, key)
+        upload_non_seekable_stream_via_tempfile(
+            client=s3,
+            bucket=bucket,
+            key=key,
+            stream=pipeline_stdout,
+            part_size=part_size,
+        )
     finally:
         # Allow the pipeline to fail if the upload fails
         pipeline_stdout.close()
diff --git a/src/btrfs2s3/stream_uploader.py b/src/btrfs2s3/stream_uploader.py
new file mode 100644
index 0000000..913045b
--- /dev/null
+++ b/src/btrfs2s3/stream_uploader.py
@@ -0,0 +1,127 @@
+"""Functions for uploading a non-seekable stream to S3."""
+
+from __future__ import annotations
+
+import os
+from tempfile import TemporaryFile
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from typing import AnyStr
+    from typing import IO
+    from typing import Iterator
+
+    from mypy_boto3_s3.client import S3Client
+    from mypy_boto3_s3.type_defs import CompletedPartTypeDef
+
+_COPY_BUFFER_SIZE = 2**20
+
+
+def _copy(*, input_file: IO[bytes], output_file: IO[bytes], count: int) -> int:
+    # notes for fast copy:
+    # - buffered IO classes have seek position out of sync with underlying file
+    #   descriptor, mitigate this with SEEK_END
+    # - sendfile and splice. which should be used first?
+    # - use readinto and reuse buffers for slow mode
+    written = 0
+    eof = False
+    while written < count and not eof:
+        # This may result in one or multiple underlying reads
+        buf = input_file.read(min(count - written, _COPY_BUFFER_SIZE))
+        if not buf:
+            eof = True
+            break
+        # https://docs.python.org/3/library/io.html#io.RawIOBase.write
+        # indicates this loop is needed
+        offset = 0
+        while offset < len(buf):
+            offset += output_file.write(buf[offset:])
+        written += len(buf)
+    return written
+
+
+def _iter_parts_via_tempfile(stream: IO[bytes], part_size: int) -> Iterator[IO[bytes]]:
+    while True:
+        with TemporaryFile() as part_file:
+            written = _copy(input_file=stream, output_file=part_file, count=part_size)
+            if written > 0:
+                part_file.seek(0, os.SEEK_SET)
+                yield part_file
+            if written < part_size:
+                break
+
+
+def _stream_len(stream: IO[AnyStr]) -> int:
+    cur = stream.tell()
+    end = stream.seek(0, os.SEEK_END)
+    stream.seek(cur, os.SEEK_SET)
+    return end
+
+
+def upload_non_seekable_stream_via_tempfile(
+    *, stream: IO[bytes], client: S3Client, bucket: str, key: str, part_size: int
+) -> None:
+    """Upload a non-seekable stream to S3.
+
+    This will store the stream in parts to temporary files, of part_size bytes
+    each. If less than one full part is consumed from the stream, it will
+    upload the object with put_object. Otherwise, a multipart upload will be
+    used.
+
+    If any error is raised, this function will attempt to cancel the multipart
+    upload with abort_multipart_upload().
+
+    Args:
+        stream: A stream to upload. The stream may be seekable, but this
+            function is designed for the non-seekable case.
+        client: The S3 client object.
+        bucket: The name of the S3 bucket.
+        key: The key of the S3 object in the bucket.
+        part_size: The maximum size of a single part.
+    """
+    # If the first part is the maximum part size, assume there will be more parts. This
+    # is suboptimal in the rare case that the stream is exactly one part length long.
+    # The alternative is to attempt to read an extra byte from the stream after the
+    # first part has been collected, and append it to the next part. The 1-byte reads
+    # will frequently be unaligned and lead to cache thrashing. The optimal strategy
+    # would be:
+    # - Read the first full part
+    # - Read 1 test byte
+    # - Read the second part minus one byte
+    # - Read the remaining parts as normal
+    # This would be a lot of code complexity for a very rare gain.
+    upload_id: str | None = None
+    completed_parts: list[CompletedPartTypeDef] = []
+    try:
+        for part_index, part_file in enumerate(
+            _iter_parts_via_tempfile(stream, part_size)
+        ):
+            if upload_id is None and _stream_len(part_file) == part_size:
+                upload_id = client.create_multipart_upload(Bucket=bucket, Key=key)[
+                    "UploadId"
+                ]
+            if upload_id is not None:
+                part_number = part_index + 1
+                up_response = client.upload_part(
+                    Bucket=bucket,
+                    Key=key,
+                    PartNumber=part_number,
+                    UploadId=upload_id,
+                    Body=part_file,
+                )
+                completed_parts.append(
+                    {"ETag": up_response["ETag"], "PartNumber": part_number}
+                )
+            else:
+                client.put_object(Bucket=bucket, Key=key, Body=part_file)
+        if upload_id is not None:
+            client.complete_multipart_upload(
+                Bucket=bucket,
+                Key=key,
+                UploadId=upload_id,
+                MultipartUpload={"Parts": completed_parts},
+            )
+    except Exception:
+        if upload_id is not None:
+            client.abort_multipart_upload(Bucket=bucket, Key=key, UploadId=upload_id)
+        raise
diff --git a/tests/stream_uploader/__init__.py b/tests/stream_uploader/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/tests/stream_uploader/conftest.py b/tests/stream_uploader/conftest.py
new file mode 100644
index 0000000..f891bac
--- /dev/null
+++ b/tests/stream_uploader/conftest.py
@@ -0,0 +1,59 @@
+from __future__ import annotations
+
+import os
+from tempfile import TemporaryFile
+from threading import Thread
+from typing import cast
+from typing import TYPE_CHECKING
+
+import pytest
+
+if TYPE_CHECKING:
+    from typing import IO
+    from typing import Iterator
+
+
+@pytest.fixture(params=[5 * 2**20, 5 * 2**20 + 512, 10 * 2**20])
+def data_size(request: pytest.FixtureRequest) -> int:
+    return cast(int, request.param)
+
+
+@pytest.fixture(params=[5 * 2**20, 5 * 2**20 + 512, 10 * 2**20])
+def part_size(request: pytest.FixtureRequest) -> int:
+    return cast(int, request.param)
+
+
+@pytest.fixture()
+def stream_data(data_size: int) -> bytes:
+    return os.urandom(data_size)
+
+
+@pytest.fixture(params=[-1, 0], ids=["buffered", "unbuffered"])
+def buffering(request: pytest.FixtureRequest) -> int:
+    return cast(int, request.param)
+
+
+@pytest.fixture(params=[False, True], ids=["nonseekable", "seekable"])
+def seekable(request: pytest.FixtureRequest) -> bool:
+    return cast(bool, request.param)
+
+
+@pytest.fixture()
+def stream(buffering: int, seekable: bool, stream_data: bytes) -> Iterator[IO[bytes]]:  # noqa: FBT001
+    if seekable:
+        with TemporaryFile(buffering=buffering) as stream:
+            stream.write(stream_data)
+            stream.seek(0, os.SEEK_SET)
+            yield stream
+    else:
+        read_fd, write_fd = os.pipe()
+
+        def fill(write_fd: int) -> None:
+            os.write(write_fd, stream_data)
+            os.close(write_fd)
+
+        fill_thread = Thread(target=fill, args=(write_fd,))
+        fill_thread.start()
+        with open(read_fd, mode="rb", buffering=buffering) as stream:  # noqa: PTH123
+            yield stream
+        fill_thread.join()
diff --git a/tests/stream_uploader/iter_parts_test.py b/tests/stream_uploader/iter_parts_test.py
new file mode 100644
index 0000000..2c39492
--- /dev/null
+++ b/tests/stream_uploader/iter_parts_test.py
@@ -0,0 +1,24 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+from btrfs2s3.stream_uploader import _iter_parts_via_tempfile
+
+if TYPE_CHECKING:
+    from typing import IO
+
+
+def test_iter_parts_via_tempfile(
+    stream: IO[bytes], stream_data: bytes, part_size: int
+) -> None:
+    expected_data = [
+        stream_data[i : i + part_size] for i in range(0, len(stream_data), part_size)
+    ]
+
+    got_data = []
+    for part_file in _iter_parts_via_tempfile(stream, part_size):
+        assert part_file.seekable()
+        assert part_file.tell() == 0
+        got_data.append(part_file.read())
+
+    assert got_data == expected_data
diff --git a/tests/stream_uploader/upload_non_seekable_stream_via_tempfile_test.py b/tests/stream_uploader/upload_non_seekable_stream_via_tempfile_test.py
new file mode 100644
index 0000000..97ab9c3
--- /dev/null
+++ b/tests/stream_uploader/upload_non_seekable_stream_via_tempfile_test.py
@@ -0,0 +1,49 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+from unittest.mock import Mock
+
+from btrfs2s3.stream_uploader import upload_non_seekable_stream_via_tempfile
+import pytest
+
+if TYPE_CHECKING:
+    from typing import IO
+
+    from mypy_boto3_s3.client import S3Client
+
+
+def test_file_is_uploaded(
+    s3: S3Client, stream: IO[bytes], bucket: str, part_size: int, stream_data: bytes
+) -> None:
+    key = "test-key"
+
+    upload_non_seekable_stream_via_tempfile(
+        client=s3, stream=stream, bucket=bucket, key=key, part_size=part_size
+    )
+
+    assert s3.get_object(Bucket=bucket, Key=key)["Body"].read() == stream_data
+
+
+class FakeError(Exception):
+    pass
+
+
+def test_multipart_upload_gets_cancelled_on_error(
+    s3: S3Client, stream: IO[bytes], bucket: str, part_size: int
+) -> None:
+    key = "test-key"
+
+    mock_client = Mock(wraps=s3)
+    mock_client.put_object.side_effect = FakeError()
+    mock_client.upload_part.side_effect = FakeError()
+
+    with pytest.raises(FakeError):
+        upload_non_seekable_stream_via_tempfile(
+            client=mock_client,
+            stream=stream,
+            bucket=bucket,
+            key=key,
+            part_size=part_size,
+        )
+
+    assert s3.list_multipart_uploads(Bucket=bucket).get("Uploads", []) == []