Skip to content

Commit

Permalink
feat(ingest): Created a faster ingestion mode - pipeline (#1750)
Browse files Browse the repository at this point in the history
* Unify pgvector and postgres connection settings

* Remove local changes

* Update file pgvector->postgres

* postgresql should be postgres

* Adding pipeline ingestion mode

* disable hugging face parallelism.  Continue on file to doc transform failure

* Semaphore to limit docq async workers. ETA reporting
  • Loading branch information
dbzoo committed Mar 19, 2024
1 parent 1efac6a commit 134fc54
Show file tree
Hide file tree
Showing 5 changed files with 301 additions and 2 deletions.
1 change: 1 addition & 0 deletions fern/docs/pages/manual/ingestion.mdx
Expand Up @@ -62,6 +62,7 @@ The following ingestion mode exist:
* `simple`: historic behavior, ingest one document at a time, sequentially
* `batch`: read, parse, and embed multiple documents using batches (batch read, and then batch parse, and then batch embed)
* `parallel`: read, parse, and embed multiple documents in parallel. This is the fastest ingestion mode for local setup.
* `pipeline`: Alternative to parallel.
To change the ingestion mode, you can use the `embedding.ingest_mode` configuration value. The default value is `simple`.

To configure the number of workers used for parallel or batched ingestion, you can use
Expand Down
175 changes: 174 additions & 1 deletion private_gpt/components/ingest/ingest_component.py
Expand Up @@ -6,19 +6,21 @@
import os
import threading
from pathlib import Path
from queue import Queue
from typing import Any

from llama_index.core.data_structs import IndexDict
from llama_index.core.embeddings.utils import EmbedType
from llama_index.core.indices import VectorStoreIndex, load_index_from_storage
from llama_index.core.indices.base import BaseIndex
from llama_index.core.ingestion import run_transformations
from llama_index.core.schema import Document, TransformComponent
from llama_index.core.schema import BaseNode, Document, TransformComponent
from llama_index.core.storage import StorageContext

from private_gpt.components.ingest.ingest_helper import IngestionHelper
from private_gpt.paths import local_data_path
from private_gpt.settings.settings import Settings
from private_gpt.utils.eta import eta

logger = logging.getLogger(__name__)

Expand Down Expand Up @@ -314,6 +316,170 @@ def __del__(self) -> None:
self._file_to_documents_work_pool.terminate()


class PipelineIngestComponent(BaseIngestComponentWithIndex):
"""Pipeline ingestion - keeping the embedding worker pool as busy as possible.
This class implements a threaded ingestion pipeline, which comprises two threads
and two queues. The primary thread is responsible for reading and parsing files
into documents. These documents are then placed into a queue, which is
distributed to a pool of worker processes for embedding computation. After
embedding, the documents are transferred to another queue where they are
accumulated until a threshold is reached. Upon reaching this threshold, the
accumulated documents are flushed to the document store, index, and vector
store.
Exception handling ensures robustness against erroneous files. However, in the
pipelined design, one error can lead to the discarding of multiple files. Any
discarded files will be reported.
"""

NODE_FLUSH_COUNT = 5000 # Save the index every # nodes.

def __init__(
self,
storage_context: StorageContext,
embed_model: EmbedType,
transformations: list[TransformComponent],
count_workers: int,
*args: Any,
**kwargs: Any,
) -> None:
super().__init__(storage_context, embed_model, transformations, *args, **kwargs)
self.count_workers = count_workers
assert (
len(self.transformations) >= 2
), "Embeddings must be in the transformations"
assert count_workers > 0, "count_workers must be > 0"
self.count_workers = count_workers
# We are doing our own multiprocessing
# To do not collide with the multiprocessing of huggingface, we disable it
os.environ["TOKENIZERS_PARALLELISM"] = "false"

# doc_q stores parsed files as Document chunks.
# Using a shallow queue causes the filesystem parser to block
# when it reaches capacity. This ensures it doesn't outpace the
# computationally intensive embeddings phase, avoiding unnecessary
# memory consumption. The semaphore is used to bound the async worker
# embedding computations to cause the doc Q to fill and block.
self.doc_semaphore = multiprocessing.Semaphore(
self.count_workers
) # limit the doc queue to # items.
self.doc_q: Queue[tuple[str, str | None, list[Document] | None]] = Queue(20)
# node_q stores documents parsed into nodes (embeddings).
# Larger queue size so we don't block the embedding workers during a slow
# index update.
self.node_q: Queue[
tuple[str, str | None, list[Document] | None, list[BaseNode] | None]
] = Queue(40)
threading.Thread(target=self._doc_to_node, daemon=True).start()
threading.Thread(target=self._write_nodes, daemon=True).start()

def _doc_to_node(self) -> None:
# Parse documents into nodes
with multiprocessing.pool.ThreadPool(processes=self.count_workers) as pool:
while True:
try:
cmd, file_name, documents = self.doc_q.get(
block=True
) # Documents for a file
if cmd == "process":
# Push CPU/GPU embedding work to the worker pool
# Acquire semaphore to control access to worker pool
self.doc_semaphore.acquire()
pool.apply_async(
self._doc_to_node_worker, (file_name, documents)
)
elif cmd == "quit":
break
finally:
if cmd != "process":
self.doc_q.task_done() # unblock Q joins

def _doc_to_node_worker(self, file_name: str, documents: list[Document]) -> None:
# CPU/GPU intensive work in its own process
try:
nodes = run_transformations(
documents, # type: ignore[arg-type]
self.transformations,
show_progress=self.show_progress,
)
self.node_q.put(("process", file_name, documents, nodes))
finally:
self.doc_semaphore.release()
self.doc_q.task_done() # unblock Q joins

def _save_docs(
self, files: list[str], documents: list[Document], nodes: list[BaseNode]
) -> None:
try:
logger.info(
f"Saving {len(files)} files ({len(documents)} documents / {len(nodes)} nodes)"
)
self._index.insert_nodes(nodes)
for document in documents:
self._index.docstore.set_document_hash(
document.get_doc_id(), document.hash
)
self._save_index()
except Exception:
# Tell the user so they can investigate these files
logger.exception(f"Processing files {files}")
finally:
# Clearing work, even on exception, maintains a clean state.
nodes.clear()
documents.clear()
files.clear()

def _write_nodes(self) -> None:
# Save nodes to index. I/O intensive.
node_stack: list[BaseNode] = []
doc_stack: list[Document] = []
file_stack: list[str] = []
while True:
try:
cmd, file_name, documents, nodes = self.node_q.get(block=True)
if cmd in ("flush", "quit"):
if file_stack:
self._save_docs(file_stack, doc_stack, node_stack)
if cmd == "quit":
break
elif cmd == "process":
node_stack.extend(nodes) # type: ignore[arg-type]
doc_stack.extend(documents) # type: ignore[arg-type]
file_stack.append(file_name) # type: ignore[arg-type]
# Constant saving is heavy on I/O - accumulate to a threshold
if len(node_stack) >= self.NODE_FLUSH_COUNT:
self._save_docs(file_stack, doc_stack, node_stack)
finally:
self.node_q.task_done()

def _flush(self) -> None:
self.doc_q.put(("flush", None, None))
self.doc_q.join()
self.node_q.put(("flush", None, None, None))
self.node_q.join()

def ingest(self, file_name: str, file_data: Path) -> list[Document]:
documents = IngestionHelper.transform_file_into_documents(file_name, file_data)
self.doc_q.put(("process", file_name, documents))
self._flush()
return documents

def bulk_ingest(self, files: list[tuple[str, Path]]) -> list[Document]:
docs = []
for file_name, file_data in eta(files):
try:
documents = IngestionHelper.transform_file_into_documents(
file_name, file_data
)
self.doc_q.put(("process", file_name, documents))
docs.extend(documents)
except Exception:
logger.exception(f"Skipping {file_data.name}")
self._flush()
return docs


def get_ingestion_component(
storage_context: StorageContext,
embed_model: EmbedType,
Expand All @@ -336,6 +502,13 @@ def get_ingestion_component(
transformations=transformations,
count_workers=settings.embedding.count_workers,
)
elif ingest_mode == "pipeline":
return PipelineIngestComponent(
storage_context=storage_context,
embed_model=embed_model,
transformations=transformations,
count_workers=settings.embedding.count_workers,
)
else:
return SimpleIngestComponent(
storage_context=storage_context,
Expand Down
4 changes: 3 additions & 1 deletion private_gpt/settings/settings.py
Expand Up @@ -155,13 +155,14 @@ class HuggingFaceSettings(BaseModel):

class EmbeddingSettings(BaseModel):
mode: Literal["huggingface", "openai", "azopenai", "sagemaker", "ollama", "mock"]
ingest_mode: Literal["simple", "batch", "parallel"] = Field(
ingest_mode: Literal["simple", "batch", "parallel", "pipeline"] = Field(
"simple",
description=(
"The ingest mode to use for the embedding engine:\n"
"If `simple` - ingest files sequentially and one by one. It is the historic behaviour.\n"
"If `batch` - if multiple files, parse all the files in parallel, "
"and send them in batch to the embedding model.\n"
"In `pipeline` - The Embedding engine is kept as busy as possible\n"
"If `parallel` - parse the files in parallel using multiple cores, and embedd them in parallel.\n"
"`parallel` is the fastest mode for local setup, as it parallelize IO RW in the index.\n"
"For modes that leverage parallelization, you can specify the number of "
Expand All @@ -174,6 +175,7 @@ class EmbeddingSettings(BaseModel):
"The number of workers to use for file ingestion.\n"
"In `batch` mode, this is the number of workers used to parse the files.\n"
"In `parallel` mode, this is the number of workers used to parse the files and embed them.\n"
"In `pipeline` mode, this is the number of workers that can perform embeddings.\n"
"This is only used if `ingest_mode` is not `simple`.\n"
"Do not go too high with this number, as it might cause memory issues. (especially in `parallel` mode)\n"
"Do not set it higher than your number of threads of your CPU."
Expand Down
122 changes: 122 additions & 0 deletions private_gpt/utils/eta.py
@@ -0,0 +1,122 @@
import datetime
import logging
import math
import time
from collections import deque
from typing import Any

logger = logging.getLogger(__name__)


def human_time(*args: Any, **kwargs: Any) -> str:
def timedelta_total_seconds(timedelta: datetime.timedelta) -> float:
return (
timedelta.microseconds
+ 0.0
+ (timedelta.seconds + timedelta.days * 24 * 3600) * 10**6
) / 10**6

secs = float(timedelta_total_seconds(datetime.timedelta(*args, **kwargs)))
# We want (ms) precision below 2 seconds
if secs < 2:
return f"{secs * 1000}ms"
units = [("y", 86400 * 365), ("d", 86400), ("h", 3600), ("m", 60), ("s", 1)]
parts = []
for unit, mul in units:
if secs / mul >= 1 or mul == 1:
if mul > 1:
n = int(math.floor(secs / mul))
secs -= n * mul
else:
# >2s we drop the (ms) component.
n = int(secs)
if n:
parts.append(f"{n}{unit}")
return " ".join(parts)


def eta(iterator: list[Any]) -> Any:
"""Report an ETA after 30s and every 60s thereafter."""
total = len(iterator)
_eta = ETA(total)
_eta.needReport(30)
for processed, data in enumerate(iterator, start=1):
yield data
_eta.update(processed)
if _eta.needReport(60):
logger.info(f"{processed}/{total} - ETA {_eta.human_time()}")


class ETA:
"""Predict how long something will take to complete."""

def __init__(self, total: int):
self.total: int = total # Total expected records.
self.rate: float = 0.0 # per second
self._timing_data: deque[tuple[float, int]] = deque(maxlen=100)
self.secondsLeft: float = 0.0
self.nexttime: float = 0.0

def human_time(self) -> str:
if self._calc():
return f"{human_time(seconds=self.secondsLeft)} @ {int(self.rate * 60)}/min"
return "(computing)"

def update(self, count: int) -> None:
# count should be in the range 0 to self.total
assert count > 0
assert count <= self.total
self._timing_data.append((time.time(), count)) # (X,Y) for pearson

def needReport(self, whenSecs: int) -> bool:
now = time.time()
if now > self.nexttime:
self.nexttime = now + whenSecs
return True
return False

def _calc(self) -> bool:
# A sample before a prediction. Need two points to compute slope!
if len(self._timing_data) < 3:
return False

# http://en.wikipedia.org/wiki/Pearson_product-moment_correlation_coefficient
# Calculate means and standard deviations.
samples = len(self._timing_data)
# column wise sum of the timing tuples to compute their mean.
mean_x, mean_y = (
sum(i) / samples for i in zip(*self._timing_data, strict=False)
)
std_x = math.sqrt(
sum(pow(i[0] - mean_x, 2) for i in self._timing_data) / (samples - 1)
)
std_y = math.sqrt(
sum(pow(i[1] - mean_y, 2) for i in self._timing_data) / (samples - 1)
)

# Calculate coefficient.
sum_xy, sum_sq_v_x, sum_sq_v_y = 0.0, 0.0, 0
for x, y in self._timing_data:
x -= mean_x
y -= mean_y
sum_xy += x * y
sum_sq_v_x += pow(x, 2)
sum_sq_v_y += pow(y, 2)
pearson_r = sum_xy / math.sqrt(sum_sq_v_x * sum_sq_v_y)

# Calculate regression line.
# y = mx + b where m is the slope and b is the y-intercept.
m = self.rate = pearson_r * (std_y / std_x)
y = self.total
b = mean_y - m * mean_x
x = (y - b) / m

# Calculate fitted line (transformed/shifted regression line horizontally).
fitted_b = self._timing_data[-1][1] - (m * self._timing_data[-1][0])
fitted_x = (y - fitted_b) / m
_, count = self._timing_data[-1] # adjust last data point progress count
adjusted_x = ((fitted_x - x) * (count / self.total)) + x
eta_epoch = adjusted_x

self.secondsLeft = max([eta_epoch - time.time(), 0])
return True
1 change: 1 addition & 0 deletions settings-local.yaml
@@ -1,3 +1,4 @@
# poetry install --extras "ui llms-llama-cpp vector-stores-qdrant embeddings-huggingface"
server:
env_name: ${APP_ENV:local}

Expand Down

0 comments on commit 134fc54

Please sign in to comment.