THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE08
63
https://www.theamericanjournals.com/index.php/tajet
PUBLISHED DATE: - 23-08-2024
https://doi.org/10.37547/tajet/Volume06Issue08-08
PAGE NO.: - 63-72
UTILIZING A SCALABLE AI/ML-BASED DATA ANOMALY
DETECTION TOOL TO IMPROVE DATA QUALITY IN
VIDEO STREAMING SERVICES
Alexander Motylev
Director, Data Test Engineering at Paramount Global / PlutoTV, Miami, Florida, United States
INTRODUCTION
Modern video streaming services, such as Netflix,
Hulu, and YouTube, play a key role in delivering
content to millions of users worldwide. With the
constant growth in data volume and the increasing
number of users, maintaining high data quality
becomes critically important for ensuring
satisfactory user experience and service reliability.
The relevance of the data quality issue lies in the
fact that poor-quality data can lead to service
disruptions, which, in turn, negatively affect users'
perception of the platform.
To timely detect anomalies, two main approaches
exist: supervised and unsupervised. Supervised
methods require labeled data, while unsupervised
methods, on the contrary, do not require labeled
data and are based on the assumption that
anomalies statistically differ from normal samples.
Additionally, deep learning methods, such as
autoencoders and recurrent neural networks
(RNNs), have shown high effectiveness in detecting
anomalies in complex and high-dimensional data.
However, their use is associated with high
computational costs and the need for significant
data to train the models. These limitations
underscore the need for developing new
approaches that are more efficient and cost-
effective.
This work aims to explore AI/ML-based tools for
anomaly detection in data to improve data quality
in video streaming services.
APPLICATION OF AI AND ML IN VIDEO
STREAMING DATA ANALYSIS
Modern streaming platforms actively integrate
artificial intelligence (AI) to optimize their
RESEARCH ARTICLE
Open Access
Abstract
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE08
64
https://www.theamericanjournals.com/index.php/tajet
operations and create new opportunities for their
audience. Currently, services like Netflix, Hulu,
Spotify, YouTube, TikTok, and many others employ
AI in their operations. AI algorithms meticulously
study users' viewing history, preferences, and
interactions with the platform to suggest the most
suitable content. The primary goal of AI
technologies is to increase the time users spend on
the platform and enhance their viewing
satisfaction[1].
Actor recognition on screen has become a popular
feature among viewers. Users can find out an
actor's name and filmography simply by pausing
the video. Developers use neural networks that
analyze movies and create tags with time points of
actors' appearances and facial outlines. Super
Resolution technology is used to improve the
quality of old films. Neural networks analyze
frames, increasing resolution and adding missing
details. This allows SD films to be converted to Full
HD and 2K films to 4K, enhancing viewers'
experience [2].
In terms of anomaly detection, it involves
identifying rare events that significantly differ from
other data. Machine learning methods are widely
used for anomaly detection and can be
implemented through:
1.
Supervised Anomaly Detection: This method
uses a labeled dataset that includes both normal
and anomalous examples. The model is trained on
this data to classify new observations. Popular
algorithms for this task include supervised neural
networks, support vector machines, and k-nearest
neighbor classifiers.
2.
Unsupervised Anomaly Detection: This
method does not require labeled data. It is based on
two assumptions: only a small percentage of data is
anomalous, and anomalies statistically differ from
normal samples. Data is clustered based on
similarity, and instances that deviate significantly
from the majority are classified as anomalies [3].
Deep learning, as a major area of artificial
intelligence, has immense potential in the field of
computer vision. This machine learning approach,
based on the principles of artificial neural
networks, allows for modeling complex functions
and processing large volumes of data. The process
of object detection and analysis using deep models
includes several key stages:
Initially, the data containing images with specified
classes and object boundaries need to be
annotated. The data may require preprocessing,
such as resizing, normalization, and augmentation.
Then, the model is trained, involving the
generation of candidate proposals for objects and
their classification. First, the model proposes
candidate objects that might be present in the
image, then classifies them and accurately localizes
the boundaries. After training, the model is tested
on a test dataset, evaluating metrics such as
detection accuracy, recall, localization precision,
and mean average precision (mAP) [4]. Below in
Figure 1 is the architecture of video classification.
Fig.1. Video classification architecture [5]
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE08
65
https://www.theamericanjournals.com/index.php/tajet
Finally, the screen can display real-time video
classification: every three seconds, a segment of the
video is classified as safe, suspicious, or criminal
activity.
PREDICTIVE ANALYTICS AND VIEWERSHIP
DATA FORECASTING
Predictive analytics, often referred to as predictive
analysis, is a method for forecasting future events
based on the analysis of historical data. The process
of predictive analytics relies on processing large
datasets that traditional analysis methods cannot
effectively handle [6]. Predicting user behavior in
video streaming involves several key stages:
In the first stage, data is collected about users and
their interactions with the platform. This can
include data on views, likes, comments, watch time,
visit frequency, etc. This data is gathered from
server logs, analytical systems, and other sources.
Data preprocessing involves cleaning the data from
noise
and
anomalies,
normalizing,
and
transforming it into suitable formats for further
analysis. This may include eliminating missing
values, removing duplicates, and aggregating data
for analysis over different time intervals.
Next, feature engineering takes place, where
features that will be used in machine learning
models are extracted.
The subsequent stage involves selecting and
training machine learning models [7]. The main
approaches to predicting user behavior in video
streaming include the use of machine learning
models such as neural networks, recurrent neural
networks (RNNs), and graph neural networks
(GNNs). These models help analyze large volumes
of user behavior data, including viewing history,
time spent on the platform, and interactions with
various content. For instance, the STAMP (Short-
Term Attention/Memory Priority Model) is used to
predict user preferences based on short-term
memory and attention [8].
Modern methods of viewership data forecasting
are based on the use of machine learning and
statistical analysis. Classical statistical models,
such as autoregressive integrated moving averages
(ARIMA), are widely used for time series analysis.
However, with the advancement of technology and
the growth of data volume, more complex methods
such as neural networks, including recurrent
neural networks (RNNs) and long short-term
memory (LSTM) networks, have come to the
forefront [9]. ARIMA (AutoRegressive Integrated
Moving Average) models are powerful tools for
analyzing and forecasting time series. The ARIMA
model includes three main components:
- Autoregression (AR): Using dependencies
between observations and their previous values.
- Integration (I): Differencing the data to eliminate
non-stationarity.
- Moving Average (MA): Modeling the dependence
of the current value of the series on the errors of
previous forecasts [10, 11].
Machine learning (ML) is one of the most powerful
tools for viewership data forecasting. Due to its
ability to process large volumes of data and identify
patterns, ML methods significantly improve
forecast accuracy compared to traditional
statistical methods [12]. However, to achieve the
greatest efficiency, hybrid models are often used.
These models are combinations of various
methodologies, leveraging the strengths of each
applied method. The main idea is to compensate for
the weaknesses of one model by utilizing the
advantages of another [13, 14].
PRACTICAL IMPLEMENTATION OF ANOMALY
DETECTION AND FORECASTING SYSTEMS FOR
VIDEO STREAMING
Anomaly detection and forecasting in video
streaming are critical tasks that can significantly
improve the quality of service and reliability of
video streaming systems. Solving this task without
specialized software is practically impossible.
These programs must possess high flexibility,
performance, and accuracy to adapt to changing
data and requirements [15].
First to consider is Apache Kafka, which has
become the de facto standard for processing
streaming data in recent years. Previously,
RabbitMQ, ActiveMQ, and other message queue
systems were used, utilizing various message
distribution patterns to distribute data between
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE08
66
https://www.theamericanjournals.com/index.php/tajet
consumers from producers. However, the scale of
these tools was limited. Kafka Connect provides a
variety of connectivity options, allowing Kafka to
integrate with any data sources.
Apache Flink stands out for its ability to process
continuous data streams on a scalable basis. Flink's
unique architecture allows it to work with both
batch and streaming data, making it a versatile tool
for various tasks. Flink's integration with Kafka is
highly seamless, ensuring smooth interaction and
supporting exactly-once delivery semantics. This
means that each event will be processed exactly
once, even in case of system failures [16].
TensorFlow, developed by Google, is one of the
most popular libraries for machine learning and
deep learning. In the context of video streaming,
TensorFlow provides flexible and powerful tools
for processing and analyzing large volumes of
video data. PyTorch, created by Facebook, is also an
essential tool for developing machine learning and
deep learning models. PyTorch supports dynamic
computational graphs, allowing for more intuitive
model development and testing. In anomaly
detection and forecasting tasks in video streaming,
PyTorch is used to create and train models capable
of analyzing time series video data and identifying
deviations. Its intuitive interface and powerful
visualization tools facilitate rapid prototyping and
model optimization.
Scikit-learn is a machine-learning library for
Python that focuses on simplicity and efficiency.
While scikit-learn is not designed for deep learning,
it offers a wide range of algorithms and tools for
traditional machine learning that can be useful for
anomaly detection in video streaming. For
instance, clustering algorithms such as K-means
and outlier detection methods such as Isolation
Forest can be used to analyze video data and detect
anomalies. Scikit-learn also provides tools for data
preprocessing and model evaluation, making it
easier to integrate these components into video
streaming systems [17].
Prophet, developed by Facebook, is one of the
leading tools for time series forecasting. Prophet is
designed with an emphasis on ease of use and high
prediction accuracy. It excels particularly with time
series that contain seasonal variations and trends,
making it ideal for video streaming analysis where
such patterns may be present. Prophet uses
additive models for forecasting, which can account
for various components of the time series, such as
seasonality, trends, and holidays. This capability
allows for effectively modeling complex time series
and accurately predicting future values.
StatsModels is another powerful tool for time
series
analysis
and
statistical
modeling.
StatsModels offers a wide range of capabilities for
statistical analysis and model building, including
methods for working with time series such as
autoregressive
integrated
moving
average
(ARIMA), seasonal autoregressive integrated
moving average (SARIMA), and exponential
smoothing. These models enable detailed analysis
of time series and accurate forecasts based on
historical data. In the context of video streaming,
this means the ability to predict server load,
potential failures, and changes in user behavior
[18].
For reliable and scalable storage, as well as for
conducting advanced data analysis, tools such as
Elasticsearch and Kibana are often used. These
tools provide powerful capabilities for indexing,
searching, and visualizing data, which is especially
important when working with large volumes of
video data. Elasticsearch is ideal for storing and
searching large volumes of data, making it
indispensable for video streaming systems that
generate vast amounts of logs and metadata.
Kibana provides a user-friendly interface for
creating dashboards and visualizing data stored in
Elasticsearch. Kibana allows users to quickly create
graphs, charts, and maps that help better
understand the system's state and identify key
trends and anomalies. Kibana's interactive
capabilities enable detailed data analysis and
sharing of results with other team members [19].
These tools allow for the automation of workflows,
efficient resource management, and high
performance, ultimately improving user service
quality and enhancing the stability of video
streaming platforms [20, 21].
Next, we will consider the process of integrating
components and data flows in video streaming
systems. In modern video streaming architectures,
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE08
67
https://www.theamericanjournals.com/index.php/tajet
components such as AI/ML models (e.g.,
TensorFlow, PyTorch, scikit-learn), predictive
analytics tools (Prophet, StatsModels), data storage
and analysis systems (Elasticsearch, Kibana), as
well as orchestration and scaling tools (Apache
Airflow, Kubernetes), and monitoring tools
(Prometheus, Grafana) play a crucial role. Proper
integration of these components allows for the
creation of an efficient and scalable video
streaming system [22]. Specific examples of
creating such a system will be presented next.
1. Customizing Prometheus and Grafana
global:
scrape_interval: 15s
scrape_configs:
- job_name: 'app_metrics'
static_configs:
- targets: ['localhost:8000']
Docker Compose for Prometheus and Grafana (docker-compose.yml):
version: '3.7'
services:
prometheus:
image: prom/prometheus
volumes:
- ./prometheus.yml:/etc/prometheus/prometheus.yml
ports:
- "9090:9090"
grafana:
image: grafana/grafana
ports:
- "3000:3000"
environment:
- GF_SECURITY_ADMIN_PASSWORD=admin
volumes:
- grafana-storage:/var/lib/grafana
volumes:
grafana-storage:
2. Configuring Apache Airflow
Docker Compose for Apache Airflow (docker-compose-airflow.yml):
version: '3.7'
services:
postgres:
image: postgres:13
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE08
68
https://www.theamericanjournals.com/index.php/tajet
environment:
POSTGRES_USER: airflow
POSTGRES_PASSWORD: airflow
POSTGRES_DB: airflow
webserver:
image: apache/airflow:2.1.0
environment:
AIRFLOW__CORE__SQL_ALCHEMY_CONN:
postgresql+psycopg2://airflow:airflow@postgres/airflow
AIRFLOW__CORE__EXECUTOR: LocalExecutor
volumes:
- ./dags:/usr/local/airflow/dags
ports:
- "8080:8080"
depends_on:
- postgres
command: webserver
scheduler:
image: apache/airflow:2.1.0
environment:
AIRFLOW__CORE__SQL_ALCHEMY_CONN:
postgresql+psycopg2://airflow:airflow@postgres/airflow
AIRFLOW__CORE__EXECUTOR: LocalExecutor
volumes:
- ./dags:/usr/local/airflow/dags
depends_on:
- postgres
command: scheduler
3. Example DAG for Apache Airflow
ETL DAG (dags/etl_dag.py):
from airflow import DAG
from airflow.operators.python_operator import PythonOperator
from airflow.utils.dates import days_ago
import pandas as pd
def extract_data(**kwargs):
# Пример извлечения данных
data = {'time': ['2024-07-17 10:00:00', '2024-07-17 10:01:00'],
'cpu_usage': [20, 30]}
df = pd.DataFrame(data)
df.to_csv('/tmp/data.csv', index=False)
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE08
69
https://www.theamericanjournals.com/index.php/tajet
def transform_data(**kwargs):
df = pd.read_csv('/tmp/data.csv')
df['cpu_usage'] = df['cpu_usage'] * 1.1 # Пример трансформации данных
df.to_csv('/tmp/data_transformed.csv', index=False)
def load_data(**kwargs):
df = pd.read_csv('/tmp/data_transformed.csv')
# Пример загрузки данных в Elasticsearch
# es = Elasticsearch(['http://localhost:9200'])
# for _, row in df.iterrows():
# es.index(index='metrics', div=row.to_dict())
default_args = {
'owner': 'airflow',
'start_date': days_ago(1),
}
with DAG(
dag_id='etl_dag',
default_args=default_args,
schedule_interval='@daily',
) as dag:
extract = PythonOperator(
task_id='extract_data',
python_callable=extract_data,
)
transform = PythonOperator(
task_id='transform_data',
python_callable=transform_data,
)
load = PythonOperator(
task_id='load_data',
python_callable=load_data,
)
extract >> transform >> load
These examples demonstrate the basic integration
of components for video streaming, including
monitoring with Prometheus and Grafana, and
orchestrating ETL processes with Apache Airflow.
Working together, they enable efficient data
management and high performance of the video
streaming system [23].
Having reviewed the examples, it is necessary to
move on to the study of implementing AI/ML
models for anomaly detection and forecasting.
There are numerous methods for anomaly
detection, which can be divided into statistical
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE08
70
https://www.theamericanjournals.com/index.php/tajet
methods, machine learning-based methods, and
deep learning-based methods. Statistical methods
include time series analysis and distribution
analysis, which allow for identifying deviations
from normal behavior. Machine learning-based
methods include clustering and classification,
where models are trained to distinguish between
normal and anomalous data. Deep learning-based
methods, such as autoencoders and recurrent
neural networks (RNNs), enable efficient
processing of complex and high-dimensional data
[24].
Deployment and scaling of the system. Initially, the
deployment process involves installing and
configuring software on target servers or in cloud
infrastructure. At this stage, it is crucial to ensure
the correct operation of all system components,
their interaction, and compliance with stated
requirements. Various automation tools, such as
Ansible, Chef, Puppet, and others, can be used for
this purpose. These tools allow standardizing the
deployment process, reducing the likelihood of
errors, and accelerating the implementation of new
software versions.
Scaling the system is aimed at ensuring its stable
operation under increasing load. Scaling can be
horizontal or vertical. Horizontal scaling involves
adding additional servers or containers, while
vertical scaling involves increasing the power of
existing servers by adding resources such as RAM
or CPU cores. Both approaches have their
advantages and disadvantages, and the choice
between them depends on specific operating
conditions and system architecture [25].
Containerization and cloud technologies represent
revolutionary
approaches
in
software
development, deployment, and operation [26].
Containerization is based on the concept of
isolating applications and their dependencies into
containers that can run independently of the
environment in which they are executed. The main
containerization tool is Docker, which allows
developers to create, test, and deploy applications
in standardized containers. Containers provide
ease of transfer and environment reproducibility,
significantly simplifying the software development
and deployment process.
Cloud technologies provide the infrastructure and
platforms
for
deploying
and
managing
containerized applications. Cloud providers, such
as Amazon Web Services (AWS), Microsoft Azure,
and Google Cloud Platform (GCP), offer a wide
range of services, including computing resources,
data storage, and container management tools such
as Kubernetes. Kubernetes, in particular, has
become the de facto standard for container
orchestration,
providing
mechanisms
for
automatic deployment, scaling, and management of
containerized applications [27].
CONCLUSION
This research examined the problem of improving
data quality in video streaming services using
scalable AI/ML-based anomaly detection tools. The
focus was on analyzing existing methods, such as
supervised and unsupervised learning, as well as
deep learning, and identifying their limitations. The
proposed solution includes the use of advanced
machine learning methods for effectively detecting
anomalies in data and predicting user behavior.
The developed anomaly detection system has
significant potential for improving data quality in
video streaming services. It not only efficiently
detects and eliminates anomalies but also predicts
user behavior, which enhances user experience
and service reliability. This, in turn, will lead to
increased user satisfaction and audience retention,
which is critically important for the success of
streaming platforms.
The practical significance of the proposed solution
lies in its ability to significantly reduce data
processing
costs
and
network
resource
management. Implementing AI/ML tools allows
automating content moderation processes and
managing video stream quality, leading to more
efficient resource utilization and reduced
operational costs. These advantages make the
proposed solution attractive to a wide range of
video streaming platforms seeking to improve
their service quality and increase competitiveness.
The future development prospects of this
technology include integration with new data
sources, such as social networks and user reviews,
allowing for even more accurate and timely
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE08
71
https://www.theamericanjournals.com/index.php/tajet
anomaly detection.
REFERENCES
1.
How AI simplifies and automates the work of
streaming services. [Electronic resource]
Access
mode:
https://www.comnews.ru/content/230555/2
023-12-07/2023-w49/1013/kak-ii-
uproschaet-i-avtomatiziruet-rabotu-
strimingovykh-servisov
(accessed
07/12/2024).
2.
Anitha Kumari K, Dharani D. A Survey on
Anomaly Detection using Unsupervised
Learning Techniques. [Electronic resource]
Access
mode:
https://www.researchgate.net/publication/3
43318335_A_Survey_on_Anomaly_Detection_u
sing_Unsupervised_Learning_Techniques
(accessed 12.07.2024).
3.
Popova I.A. Detection of anomalies in a data set
using machine learning algorithms without a
teacher isolation forest and local outlier factor
// StudNet. 2020. No.12. pp.1460-1470.
4.
Chernikov A.D. Prediction and recognition of
objects in a video stream using deep learning //
Bulletin of Science. 2023. No.9 (66). pp.209-
215.
5.
Video Analysis to Detect Suspicious Activity
Based on Deep Learning. [Electronic resource]
Access
mode:
https://dzone.com/articles/video-analysis-to-
detect-suspicious-activity-based
(accessed
12.07.2024).
6.
Kalytyuk I.S., Frantsuzova G.A., Gunko A.V. On
the issue of choosing methods for predictive
analysis of social media data // Automation
and software Engineering. 2019. No.4 (30).
pp.9-17.
7.
User behavior: how to understand and predict
user behavior using click modeling. [Electronic
resource]
Access
mode:
https://fastercapital.com/ru/content/Поведе
ние
-
пользователей
--
как
-
понять
-
и
-
спрогнозировать
-
поведение
-
пользователей
-
с
-
помощью
-
моделирования
-
кликов.html
(accessed
07/12/2024).
8.
Denis Khryashchev,Alexandru Papiu, Jiamin
Xuan, Olivia Dinica, Kyle Hubert, Huy Vo Who
Watches What: Forecasting Viewership for the
Top 100 TV Networks // Computational Data
and Social Networks. 2019. pp.163-174.
9.
Lyndon Nixon Predicting Your Future
Audience's Popular Topics to Optimize TV
Content Marketing Success // Conference:
Proceedings of the 2nd International
Workshop on AI for Smart TV Content
Production, Access and Delivery. 2020. p.5.
10.
Sizikov D.O. Time series analysis method and
its mathematical model in software //
International Scientific Research Journal. 2024.
No.3 (141) . [Electronic resource] Access mode:
https://research-journal.org/archive/3-141-
2024-march/10.23670/IRJ.2024.141.5
(accessed 07/12/2024).
11.
Python | ARIMA Model for Time Series
Forecasting. [Electronic resource] Access
mode:
https://translated.turbopages.org/proxy_u/en
-ru.ru.8f7a0cad-66981a60-3df9103b-
74722d776562/https/www.geeksforgeeks.or
g/python-arima-model-for-time-series-
forecasting / (accessed 12.07.2024).
12.
Dozhdikov Anton Valentinovich Forecasting
the results of film distribution using machine
learning // VTE. 2023. No.4. pp. 93-114.
13.
Hybrid Machine Learning Approach to
Popularity Prediction of Newly Released
Contents for Online Video Streaming Service.
[Electronic
resource]
Access
mode:
https://translated.turbopages.org/proxy_u/en
-ru.ru.859ed0ea-66981acf-db75a817-
74722d776562/https/paperswithcode.com/p
aper/hybrid-machine-learning-approach-to
(accessed 12.07.2024).
14.
Morozova V. I. Forecasting by machine learning
method // Young Scientist. 2022. No. 21 (416).
pp. 202-204.
15.
Kusakina D. S. Development and research of a
method for detecting anomalies in a video
stream based on the analysis of the context of a
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE08
72
https://www.theamericanjournals.com/index.php/tajet
scene // Young Scientist. 2023. No. 19 (466).
pp. 15-18.
16.
Building architectures for real-time data
processing using Apache Kafka, Flink and
Druid. [Electronic resource] Access mode:
https://habr.com/ru/companies/timeweb/ar
ticles/783758 / (accessed 12.07.2024).
17.
Anomaly
detection
with
TensorFlow.
[Electronic
resource]
Access
mode:
https://translated.turbopages.org/proxy_u/en
-ru.ru.41c286df-66981b41-8a53e5ad-
74722d776562/https/www.geeksforgeeks.or
g/anomaly-detection-with-tensorflow
/
(accessed 12.07.2024).
18.
Orlova I.V. Using the Prophet package in
forecasting time series // Fundamental
Research. 2021. No. 3. pp. 94-102.
19.
ELK Stack Tutorial: Get Started with
Elasticsearch, Logstash, Kibana, & Beats.
[Electronic
resource]
Access
mode:
https://translated.turbopages.org/proxy_u/en
-ru.ru.2192893c-66981b77-65adc37a-
74722d776562/https/www.tutorialspoint.co
m/elk-stack-tutorial-get-started-with-
elasticsearch-logstash-kibana-amp-beats
(accessed 12.07.2024).
20.
How to use the Kubernetes Executor in Airflow
in production. [Electronic resource] Access
mode: https://www.restack.io/docs/airflow-
kubernetes-executor (accessed 12.07.2024).
21.
Monitoring of WebRTC streams using
Prometheus and Grafana. [Electronic resource]
Access
mode:
https://habr.com/ru/companies/flashphoner
/articles/568784 / (accessed 07/12/2024).
22.
Prometheus vs Grafana: Top Differences.
[Electronic
resource]
Access
mode:
https://translated.turbopages.org/proxy_u/en
-ru.ru.973f9ca9-66981ba6-c413aef2-
74722d776562/https/www.geeksforgeeks.or
g/prometheus-vs-grafana
/
(accessed
12.07.2024).
23.
Integration of third-party video analytics.
[Electronic
resource]
Access
mode:
https://flussonic.ru/doc/integrate-external-
video-analytics / (accessed 12.07.2024).
24.
Machine Learning for Anomaly Detection.
[Electronic
resource]
Access
mode:
https://translated.turbopages.org/proxy_u/en
-ru.ru.7c9cebae-66981bc9-0aab5e33-
74722d776562/https/www.geeksforgeeks.or
g/machine-learning-for-anomaly-detection /
(accessed 12.07.2024).
25.
Kubernetes for DevOps. [Electronic resource]
Access
mode:
https://github.com/rustamgarifulin/books/bl
ob/master/files/Kubernetes_%D0%B4%D0%
BB%D1%8F_DevOps_%D1%80%D0%B0%D0
%B7%D0%B2%D0%B5%D1%80%D1%82%
D1%8B%D0%B2%D0%B0%D0%BD%D0%B
8%D0%B5%2C_%D0%B7%D0%B0%D0%BF
%D1%83%D1%81%D0%BA_%D0%B8_%D0
%BC%D0%B0%D1%81%D1%88%D1%82%
D0%B0%D0%B1%D0%B8%D1%80%D0%BE
%D0%B2%D0%B0%D0%BD%D0%B8%D0%
B5_%D0%B2.pdf (accessed 12.07.2024).
26.
Immersion in containerization. [Electronic
resource]
Access
mode:
https://habr.com/ru/companies/otus/article
s/767884 / (accessed 12.07.2024).
27.
CI/CD for Machine Learning: What it is &
Benefits in 2024. [Electronic resource] Access
mode: https://research.aimultiple.com/ci-cd-
machine-learning / (accessed 12.07.2024).
