Last Modified: February 4th, 2026

By

Brent Sakata

Title: Platform Selection for Autocomplete Functionality in IBP Orders

Status: Proposed

IBP Orders requires autocomplete functionality across multiple use cases within the UI for various entities such as orders and users based on different attributes. Currently, IBP does not have a platform-level search solution to support these requirements.

Key Business Drivers:

• User Experience: Best practices recommend p99 response times < 200ms for autocomplete. Nielsen group studies show that 0.1 seconds creates the feeling of instantaneous response—critical for direct manipulation UI patterns.
• Feature Requirements: The solution must support:
  • Wildcard/contains searching (prefix searching alone is insufficient)
  • Multiple entity indexes (Orders and Recipients/Users)
  • Complex search queries with filtering, grouping, and sorting capabilities
  • High performance at scale

Technical Challenges:

• Database-level solutions (PostgreSQL pg_trgm, MySQL n-gram) are CPU-intensive and do not scale adequately for the required performance targets and search requirements
• Existing database infrastructure cannot reliably meet p99 < 50ms server response time targets needed to achieve the 200ms end-to-end goal

We will deploy self-hosted Typesense as the autocomplete platform for IBP Orders, running on AWS ECS with a High-Availability (HA) cluster configuration. (Pre prod will be a single task)

Infrastructure

graph TB  
    subgraph "Client Layer"  
        User[👤 End Users]  
        CF[☁️ Cloudflare CDN]  
        Angular[🅰️ Angular App]  
    end  
     
    subgraph "AWS Cloud - us-east-1"  
        subgraph "Edge/API Layer"  
            APIGW[🚪 API Gateway<br/>REST API]  
            LambdaAuth[λ IBP Authorizer<br/>Token Validation<br/>User Identity]  
        end  
         
        subgraph "Processing Layer"  
            LambdaProxy[λ Proxy Function<br/>Add Tenant Filters<br/>Query Transformation]  
        end  
         
        subgraph "VPC - [10.0.0.0/16](http://10.0.0.0/16)"  
            subgraph "Load Balancing"  
                ALB[⚖️ Application Load Balancer<br/>Internal<br/>Health Checks]  
            end  
             
            subgraph "Availability Zone 1a"  
                ECSTask1[🐳 ECS Fargate Task 1<br/>Typesense Node<br/>1 vCPU, 4 GB RAM<br/>Leader/Follower]  
                EFS1[📁 EFS Mount]  
            end  
             
            subgraph "Availability Zone 1b"  
                ECSTask2[🐳 ECS Fargate Task 2<br/>Typesense Node<br/>1 vCPU, 4 GB RAM<br/>Follower]  
                EFS2[📁 EFS Mount]  
            end  
             
            subgraph "Availability Zone 1c"  
                ECSTask3[🐳 ECS Fargate Task 3<br/>Typesense Node<br/>1 vCPU, 4 GB RAM<br/>Follower]  
                EFS3[📁 EFS Mount]  
            end  
             
            EFSStorage[(💾 EFS File System<br/>Shared Storage<br/>Multi-AZ)]  
        end  
         
        subgraph "Supporting Services"  
            CloudWatch[📊 CloudWatch<br/>Logs & Metrics]  
            Secrets[🔐 Secrets Manager<br/>API Keys]  
            VPCLink[🔗 VPC Link<br/>Private Integration]  
        end  
    end  
     
    subgraph "Raft Cluster Communication"  
        Raft[⚡ Raft Consensus Protocol<br/>Port 8107<br/>Leader Election & Replication]  
    end  
     
    %% User Flow  
    User -->|HTTPS| CF  
    CF -->|Cached Assets| Angular  
    Angular -->|API Requests<br/>+ Auth Token| APIGW  
     
    %% API Gateway Flow  
    APIGW -->|Authorize Request| LambdaAuth  
    LambdaAuth -->|Validated<br/>User Context| APIGW  
    APIGW -->|Forward Request<br/>+ User Info| LambdaProxy  
     
    %% Lambda Proxy Flow  
    LambdaProxy -->|Add tenant_id filter<br/>Transform query| VPCLink  
    VPCLink -->|Private Network| ALB  
     
    %% Load Balancer Flow  
    ALB -->|Round Robin<br/>Health Check| ECSTask1  
    ALB -->|Round Robin<br/>Health Check| ECSTask2  
    ALB -->|Round Robin<br/>Health Check| ECSTask3  
     
    %% EFS Storage  
    ECSTask1 -.->|Mount /data| EFS1  
    ECSTask2 -.->|Mount /data| EFS2  
    ECSTask3 -.->|Mount /data| EFS3  
    EFS1 -.->|Multi-AZ Replication| EFSStorage  
    EFS2 -.->|Multi-AZ Replication| EFSStorage  
    EFS3 -.->|Multi-AZ Replication| EFSStorage  
     
    %% Raft Communication  
    ECSTask1 <-->|Raft Peering<br/>8107| Raft  
    ECSTask2 <-->|Raft Peering<br/>8107| Raft  
    ECSTask3 <-->|Raft Peering<br/>8107| Raft  
     
    %% Supporting Services
    LambdaProxy -.->|Get Config| Secrets  
    ECSTask1 -.->|Send Logs/Metrics| CloudWatch  
    ECSTask2 -.->|Send Logs/Metrics| CloudWatch  
    ECSTask3 -.->|Send Logs/Metrics| CloudWatch  
    ALB -.->|Send Metrics| CloudWatch  
     
    %% Response Flow (dotted lines for clarity)  
    ECSTask1 -.->|Search Results| ALB  
    ECSTask2 -.->|Search Results| ALB  
    ECSTask3 -.->|Search Results| ALB  
    ALB -.->|Response| LambdaProxy  
    LambdaProxy -.->|Filtered Results| APIGW  
    APIGW -.->|JSON Response| Angular  
     
    %% Styling  
    classDef userLayer fill:#e1f5ff,stroke:#01579b,stroke-width:2px  
    classDef apiLayer fill:#fff3e0,stroke:#e65100,stroke-width:2px  
    classDef computeLayer fill:#f3e5f5,stroke:#4a148c,stroke-width:2px  
    classDef storageLayer fill:#e8f5e9,stroke:#1b5e20,stroke-width:2px  
    classDef networkLayer fill:#fce4ec,stroke:#880e4f,stroke-width:2px  
    classDef supportLayer fill:#f5f5f5,stroke:#424242,stroke-width:2px  
     
    class User,CF,Angular userLayer
    class APIGW,LambdaAuth apiLayer
    class LambdaProxy,ECSTask1,ECSTask2,ECSTask3 computeLayer
    class EFSStorage,EFS1,EFS2,EFS3 storageLayer
    class ALB,VPCLink networkLayer
    class CloudWatch,Secrets,Raft supportLayer

Autocomplete Flow

sequenceDiagram  
    actor User  
    participant CF as Cloudflare<br/>(Angular App)  
    participant APIGW as API Gateway  
    participant Auth as Lambda Authorizer  
    participant Proxy as Lambda Proxy  
    participant ECS as ECS Fargate<br/>(Typesense)  
     
    User->>CF: Types in search box  
    CF->>CF: Debounce input (300ms)  
    CF->>APIGW: POST /search<br/>Authorization: Bearer {token}<br/>Body: {query: "prod"}  
     
    APIGW->>Auth: Validate token  
    Auth->>Auth: Verify JWT<br/>Extract user_id, tenant_id  
    Auth-->>APIGW: Auth context<br/>{user_id, tenant_id, role}  
     
    alt Authentication Failed  
        APIGW-->>CF: 401 Unauthorized  
        CF-->>User: Show error  
    else Authentication Successful  
        APIGW->>Proxy: Forward request<br/>+ auth context  
         
        Proxy->>Proxy: Build Typesense query<br/>filter_by: tenant_id=123  
         
        Proxy->>ECS: POST /collections/products/documents/search<br/>X-TYPESENSE-API-KEY: {admin_key}<br/>Body: {q: "prod", filter_by: "tenant_id:=123"}  
         
        ECS->>ECS: Search index with filter  
        ECS-->>Proxy: Search results<br/>[{id: 1, name: "Product A"}]  
         
        Proxy-->>APIGW: 200 OK<br/>Filtered results  
        APIGW-->>CF: Results  
        CF-->>User: Display suggestions  
    end

This decision prioritizes long-term cost efficiency, operational control, and alignment with existing AWS infrastructure while maintaining excellent performance.

1. Performance Excellence

• Self-hosted Typesense delivers p99 latency of 20-50ms (same AZ) for wildcard searches—well below the 50ms server-side target
• Supports all required query types: prefix, contains, fuzzy matching, and complex filtering
• Performance scales linearly with infrastructure; same latency characteristics as cloud version
• Cache hit rates of 70-85% on typical autocomplete patterns further improve user experience
• HA cluster eliminates single points of failure for internal tool reliability

2. Exceptional Cost Optimization

Supports 3.5M records efficiently with predictable costs for foreseeable growth to 10M+ records

3. Infrastructure Alignment

• Leverages existing AWS account, VPC, and IAM infrastructure
• ECS deployment aligns with current DevOps tooling and CI/CD pipelines
• Data remains in-house within controlled AWS environment
• Direct integration with existing observability stack (CloudWatch, DataDog, etc.)

4. Operational Control

• Full control over performance tuning, ranking algorithms, and index configuration
• Can implement custom analyzers and filters specific to IBP Orders domain
• Data remains within organizational control—no third-party dependency
• Direct access to all operational metrics and logs for debugging
• Ability to scale independently of SaaS provider tiers

5. Feature Set Alignment

• Provides typo tolerance (2 typos), faceted search, complex filtering, and sorting—all required capabilities
• Supports all IBP Order use cases: wildcard searching, multiple entity indexes, and complex grouping
• Open-source foundation enables custom feature development if needed
• No artificial limits on query throughput or data size

6. Future Flexibility and Scalability

• Linear scaling path: add more powerful instances or multi-region HA without architectural changes
• Open-source nature prevents vendor lock-in; can fork or migrate to OpenSearch if business needs diverge
• Skills and infrastructure investments directly benefit AWS ecosystem knowledge
• Option to implement advanced features (custom ranking, ML-based search, etc.) without SaaS limitations

7. Team Fit

• Leverages existing AWS and DevOps expertise within organization
• Infrastructure-as-Code (Terraform/CloudFormation) templates integrate with current practices
• Straightforward API comparable to cloud alternatives reduces integration complexity
• Excellent documentation and active open-source community support

1. People/Training

• Infrastructure expertise required: Team members (DevOps/SRE) need familiarity with ECS, auto-scaling, and infrastructure monitoring
• Developers need 4-8 hours for Typesense API and integration best practices (same as cloud option)
• Recommend 40-80 hours of architectural and implementation work upfront
• Plan 1-2 hours/month ongoing maintenance per DevOps engineer

2. Process Adjustments

• Infrastructure as Code: Develop Terraform/CloudFormation templates for cluster provisioning, backup, and disaster recovery
• Data Pipeline: Establish ETL process to sync Order and User entities to Typesense (daily or event-driven)
• Relevance Tuning: Initially configure relevance settings for different entity types; plan quarterly reviews based on user feedback
• Monitoring & Alerting: Integrate with CloudWatch, DataDog, or similar; set alerts for cluster health, disk space, memory usage, and query latency p99 exceeding 100ms
• Backup Strategy: Implement automated daily snapshots to S3; document recovery procedures
• Security: Manage API key rotation, network policies, and RBAC within VPC

3. Tooling

• AWS ECS: Container orchestration for Typesense deployment
• Terraform/CloudFormation: Infrastructure as Code for reproducible deployments
• Typesense JavaScript client library for backend/frontend integration
• AWS Systems Manager: Secrets management for API keys and credentials
• CloudWatch/DataDog: Monitoring, logging, and alerting
• Data sync tooling: Develop ETL using Lambda, Glue, or managed message queues (SQS/SNS)
• Optional: Add front-end autocomplete component library (e.g., instantsearch.js-compatible solutions)

4. Risks and Mitigation

Benefits:

• ✅ Superior Long-Term Cost: $200-400/month regardless of search volume; breaks even vs. Typesense Cloud after ~10 months; dramatically cheaper than Algolia ($280-830/month) at scale
• ✅ Excellent Performance: 20-50ms p99 latency consistently beats requirements; scales linearly with infrastructure investment
• ✅ Full Operational Control: Custom tuning, data sovereignty, and zero vendor lock-in; infrastructure remains within organizational control
• ✅ Scalability: Linear scaling path without tier constraints; can handle 10M+ records and 1000+ QPS with vertical scaling
• ✅ AWS Integration: Native AWS services (ECS, CloudWatch, IAM, VPC) simplify operations; leverages existing DevOps expertise
• ✅ Future Flexibility: Open-source foundation enables forking, custom development, or seamless migration to OpenSearch if needed
• ✅ No Vendor Lock-In: Skills and investments transfer directly to broader DevOps/AWS ecosystem
• ✅ Modern Features: Built-in typo tolerance, faceted search, and complex query support with option for custom enhancements

Drawbacks:

• ❌ Higher Initial Setup Time: 40-80 hours upfront for architecture, IaC, and cluster provisioning (vs. 3-6 hours for cloud)
• ❌ Operational Overhead: Requires 1-2 hours/month maintenance (monitoring, patching, scaling decisions); lower than OpenSearch but higher than cloud solutions
• ❌ Infrastructure Complexity: Must manage HA across AZs, failover, backup/recovery, security patches, and incident response
• ❌ Operational Risk: Infrastructure failures become team responsibility; requires runbook documentation and incident response training
• ❌ DevOps Expertise Required: Team must have AWS, ECS, and observability stack competency; knowledge silos create risk
• ❌ Scaling Complexity: Requires proactive monitoring and planned scaling; unexpected traffic spikes may impact latency until scaled (vs. automatic scaling in cloud)
• ❌ No Built-in Analytics: Must develop custom tracking for search patterns, zero-result queries, and user behavior
• ❌ Single Region (Initial): Cross-region latency 60-120ms; multi-region HA requires significant additional infrastructure

Define clear criteria to determine if this decision solves the intended problems:

Scaling Decision Gate (Quarterly Review):

• If p99 latency > 100ms or CPU > 80%: Upgrade to next tier (e.g., t3.medium → r6g.large)
• If search volume growth > 50% YoY: Plan vertical scaling; evaluate multi-region HA if global expansion needed
• If infrastructure costs exceed budget by > 15%: Review query patterns and optimize indexing strategy

Deploy self-hosted Typesense on AWS ECS as IBP Orders' autocomplete platform.

This decision prioritizes long-term value creation and operational control while maintaining excellent performance:

1. Superior Economics: $200-400/month infrastructure cost with minimal maintenance (fargate instance is managed by AWS)
2. Operational Control: Full transparency and customization; data remains within organizational control
3. Technical Soundness: Exceeds all performance and feature requirements (< 50ms p99 latency); scales linearly to 10M+ records
4. AWS Alignment: Leverages existing infrastructure, expertise, and tooling; no vendor lock-in
5. Opensearch Evaluation: Evaluate moving to Opensearch once IBP Search is implemented

Why benefits outweigh challenges:

• $200-300/month fixed cost is dramatically cheaper than Algolia ($280-830/month) and AWS Opensearch and provides better long-term value than Typesense Cloud
• Performance targets (< 50ms p99) are exceeded; HA configuration ensures reliability for internal tool
• DevOps overhead (1-2 hours/month) is reasonable given cost savings and organizational AWS expertise
• Open-source foundation and AWS integration enable future optimization without vendor constraints

• ✅ Infrastructure deployed and tested across 2 AZs with automated failover
• ✅ Launch Orders autocomplete with p99 < 50ms in production
• ✅ Achieve < 5% zero-result search queries after relevance tuning
• ✅ Maintain > 99.5% uptime (HA validation during testing)
• ✅ Infrastructure cost tracking within 10% of $200-400/month forecast
• ✅ Operational team reports manageable 1-2 hours/month maintenance burden
• ✅ Feature-complete delivery within 12 weeks from decision

• Performance Benchmarks: Tables 1-11 in solution_analysis.md provide detailed comparative analysis
• Typesense Documentation: https://typesense.org/docs/ and https://typesense.org/docs/guide/high-availability.html
• AWS EC2 Pricing: https://aws.amazon.com/ec2/pricing/on-demand/
• Nielsen Norman Group Response Time Study: https://www.nngroup.com/articles/website-response-times/
• Cost Comparison Models: See solution_analysis.md TCO calculations
• Operational Complexity Analysis: Detailed in Operational Complexity section of solution_analysis.md
• AWS ECS Best Practices: https://docs.aws.amazon.com/AmazonECS/latest/developerguide/
• Terraform AWS Provider: https://registry.terraform.io/providers/hashicorp/aws/