Deployment Architecture
Terms and Definitions
APM
Analyses per minute. Please note:
Analysis is a request for Quality (Liveness) or Biometry analysis using a single media.
A single analysis with multiple media counts as separate analyses in terms of APM.
Multiple analysis types on single media (two media for Biometry) count as separate analyses in terms of APM.
PoC
Proof of Concept
Node
A Node is a worker machine. Can be either a virtual or a physical machine.
HA
High availability
K8s
Kubernetes
SC
StorageClass
RWX
ReadWriteMany
Components' Description
Oz API components:
APP is the API front app that receives REST requests, performs preprocessing, and creates tasks for other API components.
Celery is the asynchronous task queue. API has the following celery queues:
Celery-default processes system-wide tasks.
Celery-maintenance processes maintenance tasks.
Celery-tfss processes analysis tasks.
Celery-resolution checks for completion of all nested analyses within a folder and changes folder status.
Celery-preview_convert creates a video preview for media.
Celery-beat is a CronJob for managing maintenance celery tasks.
Celery-Flower is a Celery metrics collector.
Celery-regula (optional) processes document analysis tasks.
Redis is a message broker and result backend for Celery.
RabbitMQ (optional) can be used as a message broker for Celery instead of Redis.
Nginx serves static media files for external HTTP(s) requests.
O2N (optional) processes the Blacklist analysis.
Statistic (optional) provides statistics' collection for Web UI.
Web UI provides the web interface.
BIO-Updater checks for models updates and downloads new models.
Oz BIO (TFSS) runs TensorFlow with AI models and makes decisions for incoming media.
The BIO-Updater and BIO components require access to the following external resources:
Deployment Scenarios
The deployment scenario depends on the workload you expect.
Use cases
Testing/Development purposes
Small installations with low number of APM
Typical usage with moderate load
High load with HA and autoscaling
Usage with cloud provider
~ APM
~ analyses per month
~ APM
analyses per month
APM
analyses per month
Environment
Docker
Docker
Kubernetes
HA
No
Partially
Yes
Pros
Requires a minimal amount of computing resources
Low complexity, so no high-qualified engineers are needed on-site
Easy to manage and support
Partially supports HA
Can be scaled up to support higher workload
HA and autoscaling
Observability and manageability
Allows high workload and can be scaled up
Cons
Suitable only for low loads, no high APM
No scaling and high-availability
API HA requires precise balancing
Higher staff qualification requirements
High staff qualification requirements
Additional infrastructure requirements
External resource requirements
PostgreSQL
For Kubernetes deployments:
K8s v1.25+
ingress-nginx
clusterIssuer
kube-metrics
Prometheus
clusterAutoscaler
PostgreSQL
Autoscaling is implemented on the basis of ClusterAutoscaler and must be supported by your infrastructure.
Small business or PoC
Please find the installation guide here: Docker.
Type of containerization: Docker,
Type of installation: Docker compose,
Autoscaling/HA: none.
Requirements
Software
Docker 19.03+,
Podman 4.4+,
Python 3.4+.
Storage
Depends on media quality, the type and number of analyses, and the required archive depth.
May be calculated as: [average media size] * 2 * [analyses per day] * [archive depth in days]. Please refer to this article for media size reference.
Each analysis request performs read and write operations on the storage. Any additional latency in these operations will impact the analysis time.
Staff qualification:
Basic knowledge of Linux and Docker.
Deployment
Single node.
Resources:
1 node,
16 CPU/32 RAM.
Two nodes.
Resources:
2 nodes,
16 CPU/32 RAM for the first node; 8 CPU/16 RAM for the second node.
Medium Load
Please find the installation guide here: Docker.
Type of containerization: Docker/Podman,
Type of installation: Docker compose,
Autoscaling/HA: manual scaling; HA is partially supported.
Requirements
Computational resources
Depending on load, you can change the number of nodes. However, for 5+ nodes, we recommend that you proceed to the High Load section.
From 2 to 4 Docker nodes (see schemes):
2 Nodes:
24 CPU/32 RAM per node.
3 Nodes:
16 CPU/24 RAM per node.
4 Nodes:
8 CPU/16 RAM for two nodes (each),
16 CPU/24 RAM for two nodes (each).
We recommend using external self-managed PostgreSQL database and NFS share.
Software
Docker 19.03+,
Podman 4.4+,
Python 3.4+.
Storage
Depends on media quality, the type and number of analyses, and the required archive depth.
May be calculated as: [average media size] * 2 * [analyses per day] * [archive depth in days]. Please refer to this article for media size reference.
Each analysis request performs read and write operations on the storage. Any additional latency in these operations will impact the analysis time.
Staff qualification:
Advanced knowledge of Linux, Docker, and Postgres.
Deployment
2 nodes:
3 nodes:
4 nodes:
High Load
Please find the installation guide here: Kubernetes.
Type of containerization: Type of containerization: Docker containers with Kubernetes orchestration,
Type of installation: Helm charts,
Autoscaling/HA: supports autoscaling; HA for most components.
Requirements
Computational resources
3-4 nodes. Depending on load, you can change the number of nodes.
16 CPU/32 RAM Nodes for the BIO pods,
8+ CPU/16+ RAM Nodes for all other workload.
We recommend using external self-managed PostgreSQL database.
Requires RWX (ReadWriteMany) StorageClass or NFS share.
Software
Docker 19.03+,
Python 3.4+.
Storage
Depends on media quality, the type and number of analyses, and the required archive depth.
May be calculated as: [average media size] * 2 * [analyses per day] * [archive depth in days]. Please refer to this article for media size reference.
Each analysis request performs read and write operations on the storage. Any additional latency in these operations will impact the analysis time.
Staff qualification:
Advanced knowledge of Linux, Docker, Kubernetes, and Postgres.
Deployment Scheme
Last updated
Was this helpful?
