Skip to content

Commit

Permalink
[ENH] Max batch size warning (#995)
Browse files Browse the repository at this point in the history
## Description of changes

*Summarize the changes made by this PR.*
 - Improvements & Bug fixes
- In v0.4.6 we introduced batching at the sqlite layer for the
embeddings_queue to maximize throughput through submit_embeddings calls.
SQlite has a max size of variables it can bind at once that is defined
at compile time. This changes makes the embeddings_queue introspect its
compile time flags to inform its max_batch_size. Max_batch_size is a new
property method added to the Producer class that indicates the maximum
batch size submit_embeddings() will accept. Right now we use the error
internal to submit_embeddings in the sqlite impl, but callers could use
this to pre-validate input and throw a message as well.
 - New functionality
	 - None

## Test plan
Added a test for the below and above batch size cases.

## Documentation Changes
We should update the usage guide to cover this caveat.
  • Loading branch information
HammadB committed Aug 17, 2023
1 parent f8186ff commit 2fcb377
Show file tree
Hide file tree
Showing 3 changed files with 71 additions and 3 deletions.
35 changes: 35 additions & 0 deletions chromadb/db/mixins/embeddings_queue.py
Original file line number Diff line number Diff line change
Expand Up @@ -68,9 +68,13 @@ def __init__(
self.callback = callback

_subscriptions: Dict[str, Set[Subscription]]
_max_batch_size: Optional[int]
# How many variables are in the insert statement for a single record
VARIABLES_PER_RECORD = 6

def __init__(self, system: System):
self._subscriptions = defaultdict(set)
self._max_batch_size = None
super().__init__(system)

@override
Expand Down Expand Up @@ -115,6 +119,15 @@ def submit_embeddings(
if len(embeddings) == 0:
return []

if len(embeddings) > self.max_batch_size:
raise ValueError(
f"""
Cannot submit more than {self.max_batch_size:,} embeddings at once.
Please submit your embeddings in batches of size
{self.max_batch_size:,} or less.
"""
)

t = Table("embeddings_queue")
insert = (
self.querybuilder()
Expand Down Expand Up @@ -208,6 +221,28 @@ def min_seqid(self) -> SeqId:
def max_seqid(self) -> SeqId:
return 2**63 - 1

@property
@override
def max_batch_size(self) -> int:
if self._max_batch_size is None:
with self.tx() as cur:
cur.execute("PRAGMA compile_options;")
compile_options = cur.fetchall()

for option in compile_options:
if "MAX_VARIABLE_NUMBER" in option[0]:
# The pragma returns a string like 'MAX_VARIABLE_NUMBER=999'
self._max_batch_size = int(option[0].split("=")[1]) // (
self.VARIABLES_PER_RECORD
)

if self._max_batch_size is None:
# This value is the default for sqlite3 versions < 3.32.0
# It is the safest value to use if we can't find the pragma for some
# reason
self._max_batch_size = 999 // self.VARIABLES_PER_RECORD
return self._max_batch_size

def _prepare_vector_encoding_metadata(
self, embedding: SubmitEmbeddingRecord
) -> Tuple[Optional[bytes], Optional[str], Optional[str]]:
Expand Down
10 changes: 9 additions & 1 deletion chromadb/ingest/__init__.py
Original file line number Diff line number Diff line change
Expand Up @@ -59,7 +59,15 @@ def submit_embeddings(
"""Add a batch of embedding records to the given topic. Returns the SeqIDs of
the records. The returned SeqIDs will be in the same order as the given
SubmitEmbeddingRecords. However, it is not guaranteed that the SeqIDs will be
processed in the same order as the given SubmitEmbeddingRecords."""
processed in the same order as the given SubmitEmbeddingRecords. If the number
of records exceeds the maximum batch size, an exception will be thrown."""
pass

@property
@abstractmethod
def max_batch_size(self) -> int:
"""Return the maximum number of records that can be submitted in a single call
to submit_embeddings."""
pass


Expand Down
29 changes: 27 additions & 2 deletions chromadb/test/ingest/test_producer_consumer.py
Original file line number Diff line number Diff line change
Expand Up @@ -98,15 +98,15 @@ def __call__(self, embeddings: Sequence[EmbeddingRecord]) -> None:
if len(self.embeddings) >= n:
event.set()

async def get(self, n: int) -> Sequence[EmbeddingRecord]:
async def get(self, n: int, timeout_secs: int = 10) -> Sequence[EmbeddingRecord]:
"Wait until at least N embeddings are available, then return all embeddings"
if len(self.embeddings) >= n:
return self.embeddings[:n]
else:
event = Event()
self.waiters.append((n, event))
# timeout so we don't hang forever on failure
await wait_for(event.wait(), 10)
await wait_for(event.wait(), timeout_secs)
return self.embeddings[:n]


Expand Down Expand Up @@ -322,3 +322,28 @@ async def test_multiple_topics_batch(
recieved = await consume_fns[i].get(total_produced + PRODUCE_BATCH_SIZE)
assert_records_match(embeddings_n[i], recieved)
total_produced += PRODUCE_BATCH_SIZE


@pytest.mark.asyncio
async def test_max_batch_size(
producer_consumer: Tuple[Producer, Consumer],
sample_embeddings: Iterator[SubmitEmbeddingRecord],
) -> None:
producer, consumer = producer_consumer
producer.reset_state()
max_batch_size = producer_consumer[0].max_batch_size
assert max_batch_size > 0

# Make sure that we can produce a batch of size max_batch_size
embeddings = [next(sample_embeddings) for _ in range(max_batch_size)]
consume_fn = CapturingConsumeFn()
consumer.subscribe("test_topic", consume_fn, start=consumer.min_seqid())
producer.submit_embeddings("test_topic", embeddings=embeddings)
received = await consume_fn.get(max_batch_size, timeout_secs=120)
assert_records_match(embeddings, received)

embeddings = [next(sample_embeddings) for _ in range(max_batch_size + 1)]
# Make sure that we can't produce a batch of size > max_batch_size
with pytest.raises(ValueError) as e:
producer.submit_embeddings("test_topic", embeddings=embeddings)
assert "Cannot submit more than" in str(e.value)

0 comments on commit 2fcb377

Please sign in to comment.