Video Streaming

Last updated Jul 24, 2026

Ingest video streams from IP cameras, RTSP feeds, YouTube, and local files — with optional AI inference — into AnyLog.

---

AnyLog can connect to video streams, record segmented clips to a blob database, and optionally run AI inference (e.g. YOLOv5 object detection) via a gRPC server.

---

Supported protocols

Protocol	Latency	Typical use
RTMP	~1–2s	Live ingest (OBS, Twitch, YouTube)
RTMPS	~1–2s	Secure RTMP over TLS
SRT	~1–2s	Professional contribution feeds over unreliable networks
RTSP	~1–5s	IP cameras and surveillance
HTTP/HTTPS / HLS	~1–5s	Standard HTTP streams, m3u8 playlists
Local file (MP4, MOV, MKV…)	0s	Stored video files on disk

---

Prerequisites

• MongoDB connected as the blob database (stores video segments)
• PostgreSQL or SQLite connected as the SQL database (stores metadata and detections)
• (Optional) A YOLOv5 inference gRPC server for object detection — see gRPC

---

Step 1 — Connect databases

# Blob database (MongoDB)
<connect dbms customers where type = mongo and ip = 127.0.0.1 and port = 27017 and user = demo and password = passwd>

# SQL database
<connect dbms customers where type = psql and ip = 127.0.0.1 and port = 5432 and user = demo and password = passwd>

Worth confirming: both connections above use the same logical dbms name (customers) for two different physical backends (Mongo for blobs, Postgres for SQL). If AnyLog treats blob storage and SQL storage as separate namespaces under one shared logical name, this is correct as written — but if not, this needs two distinct logical dbms names instead. Flagging since I can’t verify AnyLog’s actual behavior here.

---

Step 2 — Set video variables

video_url   = "https://www.youtube.com/watch?v=rnXIjl_Rzy4"   # Times Square live
video_host  = 127.0.0.1
video_port  = 8888
video_table = video_table

Sample stream URLs:

• Abbey Road London: https://www.youtube.com/watch?v=57w2gYXjRic
• Times Square: https://www.youtube.com/watch?v=rnXIjl_Rzy4

---

Step 3 — Import the display function

import function where import_name = imshow and lib = external_lib.video_processing.cv2_stream_imshow and method = init_class
set function params where import_name = imshow and param_name = port and param_type = int and param_value = !video_port
set function params where import_name = imshow and param_name = host and param_value = !video_host

---

Step 4 — Connect to the video stream

Without inference

<video connect where
  name           = youtube and
  protocol       = https and
  interface      = url and
  address        = !video_url and
  video_dbms     = customers and
  video_table    = video_table>

With object detection inference

First start the gRPC inference client (see gRPC for setup):

<run grpc client where
  name = yolov5 and ip = 127.0.0.1 and port = 50051 and
  grpc_dir = /app/AnyLog-Network/external_lib/frame_modeling and
  proto = infer and function = PredictStream and
  request = PredictRequest and response = PredictResponse and
  service = InferenceService and debug = false and invoke = true>

Then connect with detection columns:

<video connect where
  name                 = youtube and
  protocol             = https and
  interface            = url and
  address              = !video_url and
  video_dbms           = customers and
  video_table          = video_table and
  detection_dbms       = customers and
  detection_table      = detection_table and
  detection_column     = person and
  detection_column     = car and
  detection_column     = truck and
  detection_column     = bus and
  recording_segment_time = 1 and
  detection_ignore_time  = 10>

video connect parameter reference

Parameter	Description
name	Logical name for this stream connection
protocol	Stream protocol: https, rtsp, rtmp, srt, etc.
interface	url for network streams
address	Stream URL or file path
video_dbms	Blob database for storing video segments
video_table	Table name for video metadata
detection_dbms	Database for inference results
detection_table	Table for inference results
detection_column	Object class to detect (repeat for each class)
recording_segment_time	Length of each recorded clip in minutes
detection_ignore_time	Seconds to suppress duplicate detections of the same object

---

Step 5 — Start the stream

# Without inference
run video stream where name = youtube and import_display = imshow

# With inference
run video stream where name = youtube and import_display = imshow and grpc_name = yolov5

View the live stream in a browser:

http://[video_host]:[video_port]/stream/[name]

---

Step 6 — Stop the stream

exit video where name = youtube

---

Querying video data

Video segments (no inference)

run client () sql customers format=json and stat=false \
  "select file, timestamp from video_table order by timestamp DESC limit 20"

Detection results (with inference)

run client () sql customers format=json and stat=false \
  "select file, timestamp, car, truck, bus, person from detection_table order by timestamp DESC limit 20"

---

Architecture notes

The stream runs three internal threads:

Thread	Role
Capture	Reads frames from the video source
Display	Shows frames in the real-time browser window
Storage	Writes frames to disk in H.264/yuv420p segments

• Display buffer: max 2 frames (prevents lag)
• Recording buffer: max 120 frames (~4 seconds at 30fps)
• Detections are batched (25 entries per write) and deduplicated using detection_ignore_time

node_configs.env