helios-persistence

Polyglot persistence layer for the Helios FHIR Server.

Overview

Traditional FHIR server implementations force all resources into a single database technology, creating inevitable trade-offs. A patient lookup by identifier, a population cohort query, relationship traversals through care teams, and semantic similarity searches for clinical trial matching all have fundamentally different performance characteristics, yet they're typically crammed into one system optimized for none of them.

Polyglot persistence is an architectural approach where different types of data and operations are routed to the storage technologies best suited for how that data will be accessed. Rather than accepting compromise, this pattern leverages specialized storage systems optimized for specific workloads:

Workload	Optimal Technology	Why
ACID transactions	PostgreSQL	Strong consistency guarantees
Document storage	MongoDB	Natural alignment with FHIR's resource model
Relationship traversal	Neo4j	Efficient graph queries for references
Full-text search	Elasticsearch	Optimized inverted indexes
Semantic search	Vector databases	Embedding similarity for clinical matching
Bulk analytics & ML	Object Storage	Cost-effective columnar storage

Polyglot Query Example

Consider a complex clinical query that combines multiple access patterns:

GET /Observation?patient.name:contains=smith&_text=cardiac&code:below=http://loinc.org|8867-4&_include=Observation:patient

This query requires:

1. Chained search (patient.name:contains=smith) - Find observations where the referenced patient's name contains "smith"
2. Full-text search (_text=cardiac) - Search narrative text for "cardiac"
3. Terminology subsumption (code:below=LOINC|8867-4) - Find codes that are descendants of heart rate
4. Reference resolution (_include=Observation:patient) - Include the referenced Patient resources

In a polyglot architecture, the CompositeStorage routes each component to its optimal backend:

// Conceptual flow - CompositeStorage coordinates backends
async fn search(&self, query: SearchQuery) -> SearchResult {
    // 1. Route chained search to graph database (efficient traversal)
    let patient_refs = self.neo4j.find_patients_by_name("smith").await?;

    // 2. Route full-text to Elasticsearch (optimized inverted index)
    let text_matches = self.elasticsearch.text_search("cardiac").await?;

    // 3. Route terminology query to terminology service + primary store
    let code_matches = self.postgres.codes_below("8867-4").await?;

    // 4. Intersect results and fetch from primary storage
    let observation_ids = intersect(patient_refs, text_matches, code_matches);
    let observations = self.postgres.batch_read(observation_ids).await?;

    // 5. Resolve _include from primary storage
    let patients = self.postgres.resolve_references(&observations, "patient").await?;

    SearchResult { resources: observations, included: patients }
}

No single database excels at all four operations. PostgreSQL would struggle with the graph traversal, Neo4j isn't optimized for full-text search, and Elasticsearch can't efficiently handle terminology hierarchies. Polyglot persistence lets each system do what it does best.

Architecture

helios-persistence/
├── src/
│   ├── lib.rs           # Main entry point and re-exports
│   ├── error.rs         # Comprehensive error types
│   ├── tenant/          # Multitenancy support
│   │   ├── id.rs        # Hierarchical TenantId
│   │   ├── context.rs   # TenantContext (required for all operations)
│   │   ├── permissions.rs # Fine-grained TenantPermissions
│   │   └── tenancy.rs   # TenancyModel configuration
│   ├── types/           # Core domain types
│   │   ├── stored_resource.rs    # Resource with persistence metadata
│   │   ├── search_params.rs      # Full FHIR search parameter model
│   │   ├── search_capabilities.rs # Search capability reporting
│   │   └── pagination.rs         # Cursor and offset pagination
│   ├── core/            # Storage trait hierarchy
│   │   ├── backend.rs      # Backend abstraction with capabilities
│   │   ├── storage.rs      # ResourceStorage (CRUD)
│   │   ├── versioned.rs    # VersionedStorage (vread, If-Match)
│   │   ├── history.rs      # History providers (instance/type/system)
│   │   ├── search.rs       # Search providers (basic, chained, include)
│   │   ├── transaction.rs  # ACID transactions with bundle support
│   │   ├── capabilities.rs # Runtime capability discovery
│   │   ├── bulk_export.rs  # FHIR Bulk Data Export job/data traits
│   │   ├── bulk_export_output.rs # ExportOutputStore trait
│   │   ├── bulk_export_worker.rs # Bulk export worker runtime and leasing traits
│   │   └── bulk_submit.rs  # FHIR Bulk Submit traits
│   ├── search/          # Search parameter infrastructure
│   │   ├── registry.rs     # SearchParameterRegistry (in-memory cache)
│   │   ├── loader.rs       # SearchParameterLoader (R4 standard params)
│   │   ├── extractor.rs    # FHIRPath-based value extraction
│   │   ├── converters.rs   # Type conversion utilities
│   │   ├── writer.rs       # Search index writer
│   │   ├── reindex.rs      # Reindexing operations
│   │   └── errors.rs       # Search-specific error types
│   ├── strategy/        # Tenancy isolation strategies
│   │   ├── shared_schema.rs       # tenant_id column + optional RLS
│   │   ├── schema_per_tenant.rs   # PostgreSQL search_path isolation
│   │   └── database_per_tenant.rs # Complete database isolation
│   ├── backends/        # Backend implementations
│   │   ├── sqlite/      # Reference implementation (complete)
│   │   │   ├── backend.rs      # SqliteBackend with connection pooling
│   │   │   ├── storage.rs      # ResourceStorage implementation
│   │   │   ├── transaction.rs  # TransactionProvider implementation
│   │   │   ├── schema.rs       # Schema migrations (v1-v6)
│   │   │   ├── search_impl.rs  # SearchProvider implementation
│   │   │   ├── bulk_export.rs  # BulkExportStorage implementation
│   │   │   ├── bulk_submit.rs  # BulkSubmitProvider implementation
│   │   │   └── search/         # Search query building
│   │   │       ├── query_builder.rs      # SQL query construction
│   │   │       ├── chain_builder.rs      # Chained parameter resolution
│   │   │       ├── filter_parser.rs      # _filter parameter parsing
│   │   │       ├── fts.rs                # FTS5 full-text search
│   │   │       ├── modifier_handlers.rs  # Search modifier logic
│   │   │       ├── strategy.rs           # Query strategy selection
│   │   │       ├── writer.rs             # Index writing
│   │   │       └── parameter_handlers/   # Type-specific handlers
│   │   │           ├── string.rs, token.rs, date.rs, number.rs
│   │   │           ├── quantity.rs, reference.rs, uri.rs, composite.rs
│   │   ├── postgres/       # PostgreSQL primary backend
│   │   │   ├── backend.rs      # PostgresBackend with connection pooling
│   │   │   ├── storage.rs      # ResourceStorage implementation
│   │   │   ├── transaction.rs  # TransactionProvider implementation
│   │   │   ├── schema.rs       # Schema DDL with migrations
│   │   │   ├── search_impl.rs  # SearchProvider implementation
│   │   │   ├── bulk_export.rs  # BulkExportStorage implementation
│   │   │   ├── bulk_submit.rs  # BulkSubmitProvider implementation
│   │   │   └── search/         # Search query building
│   │   │       ├── query_builder.rs  # SQL with $N params, ILIKE, TIMESTAMPTZ
│   │   │       └── writer.rs        # Search index writer
│   │   ├── mongodb/       # MongoDB primary backend
│   │   │   ├── backend.rs      # MongoBackend + MongoBackendConfig
│   │   │   ├── schema.rs       # Schema/index bootstrap helpers
│   │   │   ├── search_impl.rs  # SearchProvider implementation
│   │   │   ├── storage.rs      # ResourceStorage/history/versioning implementation
│   │   │   └── mod.rs          # Module wiring and re-exports
│   │   ├── elasticsearch/  # Search-optimized secondary backend
│   │   │   ├── backend.rs      # ElasticsearchBackend with config
│   │   │   ├── storage.rs      # ResourceStorage for sync support
│   │   │   ├── schema.rs       # Index mappings and templates
│   │   │   ├── search_impl.rs  # SearchProvider, TextSearchProvider
│   │   │   └── search/         # ES Query DSL translation
│   │   │       ├── query_builder.rs      # FHIR SearchQuery → ES Query DSL
│   │   │       ├── fts.rs                # Full-text search queries
│   │   │       ├── modifier_handlers.rs  # :missing and other modifiers
│   │   │       └── parameter_handlers/   # Type-specific handlers
│   │   │           ├── string.rs, token.rs, date.rs, number.rs
│   │   │           ├── quantity.rs, reference.rs, uri.rs, composite.rs
│   │   └── s3/               # AWS S3 object-storage backend
│   │       ├── backend.rs        # S3Backend with connection management
│   │       ├── config.rs         # S3BackendConfig, S3TenancyMode
│   │       ├── client.rs         # S3Api trait and AwsS3Client implementation
│   │       ├── keyspace.rs       # S3Keyspace key-path generation
│   │       ├── models.rs         # HistoryIndexEvent, SubmissionState
│   │       ├── storage.rs        # ResourceStorage implementation
│   │       ├── bundle.rs         # Batch/transaction bundle processing
│   │       ├── bulk_export.rs    # ExportDataProvider implementation
│   │       ├── output_store.rs   # S3OutputStore for bulk export files
│   │       ├── bulk_submit.rs    # BulkSubmitProvider implementation
│   │       └── tests.rs          # Integration tests
│   ├── composite/       # Multi-backend coordination
│   │   ├── config.rs       # CompositeConfig and builder
│   │   ├── analyzer.rs     # Query feature detection
│   │   ├── router.rs       # Query routing logic
│   │   ├── cost.rs         # Cost-based optimization
│   │   ├── merger.rs       # Result merging strategies
│   │   ├── sync.rs         # Backend synchronization
│   │   ├── health.rs       # Health monitoring
│   │   └── storage.rs      # CompositeStorage implementation
│   └── advisor/         # Configuration advisor HTTP API
│       ├── server.rs       # Axum HTTP server
│       ├── handlers.rs     # API endpoint handlers
│       ├── analysis.rs     # Configuration analysis
│       ├── suggestions.rs  # Optimization suggestions
│       └── main.rs         # Advisor binary entry point
└── tests/               # Integration tests
    ├── common/          # Shared test utilities
    │   ├── harness.rs      # Test harness setup
    │   ├── fixtures.rs     # FHIR resource fixtures
    │   ├── assertions.rs   # Custom test assertions
    │   └── capabilities.rs # Capability test helpers
    ├── crud/            # CRUD operation tests
    │   ├── create_tests.rs, read_tests.rs, update_tests.rs
    │   ├── delete_tests.rs, conditional_tests.rs
    ├── search/          # Search parameter tests
    │   ├── string_tests.rs, token_tests.rs, date_tests.rs
    │   ├── number_tests.rs, quantity_tests.rs, reference_tests.rs
    │   ├── chained_tests.rs, include_tests.rs
    │   ├── modifier_tests.rs, pagination_tests.rs
    ├── versioning/      # Version history tests
    │   ├── vread_tests.rs, history_tests.rs
    │   └── optimistic_locking_tests.rs
    ├── transactions/    # Transaction tests
    │   ├── basic_tests.rs, bundle_tests.rs, rollback_tests.rs
    ├── multitenancy/    # Tenant isolation tests
    │   ├── isolation_tests.rs, cross_tenant_tests.rs
    ├── composite_routing_tests.rs   # Query routing tests
    ├── composite_polyglot_tests.rs  # Multi-backend tests
    ├── sqlite_tests.rs              # SQLite backend tests
    ├── postgres_tests.rs            # PostgreSQL backend tests
    ├── mongodb_tests.rs             # MongoDB backend tests
    └── elasticsearch_tests.rs       # Elasticsearch backend tests

Trait Hierarchy

The storage layer uses a progressive trait hierarchy inspired by Diesel:

Backend (connection management, capabilities)
    │
    ├── ResourceStorage (create, read, update, delete)
    │       │
    │       └── VersionedStorage (vread, update_with_match)
    │               │
    │               └── HistoryProvider (instance, type, system history)
    │
    ├── SearchProvider (search, search_count)
    │       │
    │       ├── IncludeProvider (_include resolution)
    │       ├── RevincludeProvider (_revinclude resolution)
    │       └── ChainedSearchProvider (chained parameters, _has)
    │
    └── TransactionProvider (begin, commit, rollback)

Features

• Multiple Backends: SQLite, PostgreSQL, Cassandra, MongoDB, Neo4j, Elasticsearch, S3
• Multitenancy: Three isolation strategies with type-level enforcement
• Full FHIR Search: All parameter types, modifiers, chaining, _include/_revinclude
• Versioning: Complete resource history with optimistic locking
• Transactions: ACID transactions with FHIR bundle support
• Capability Discovery: Runtime introspection of backend capabilities

Multitenancy

All storage operations require a TenantContext, ensuring tenant isolation at the type level. There is no way to bypass this requirement—the compiler enforces it.

Tenancy Strategies

Strategy	Isolation	Use Case
Shared Schema	tenant_id column + optional RLS	Multi-tenant SaaS with shared infrastructure
Schema-per-Tenant	PostgreSQL schemas	Logical isolation with shared database
Database-per-Tenant	Separate databases	Complete isolation for compliance

Hierarchical Tenants

use helios_persistence::tenant::TenantId;

let parent = TenantId::new("acme");
let child = TenantId::new("acme/research");
let grandchild = TenantId::new("acme/research/oncology");

assert!(child.is_descendant_of(&parent));
assert!(grandchild.is_descendant_of(&parent));
assert_eq!(grandchild.root().as_str(), "acme");

Permission Control

use helios_persistence::tenant::{TenantPermissions, Operation};

// Read-only access
let read_only = TenantPermissions::read_only();

// Custom permissions with compartment restrictions
let custom = TenantPermissions::builder()
    .allow_operations(vec![Operation::Read, Operation::Search])
    .allow_resource_types(vec!["Patient", "Observation"])
    .restrict_to_compartment("Patient", "123")
    .build();

Search

Build search queries with full FHIR search support:

use helios_persistence::types::{
    SearchQuery, SearchParameter, SearchParamType, SearchValue,
    SearchModifier, SortDirective, IncludeDirective, IncludeType,
};

// Simple search
let query = SearchQuery::new("Patient")
    .with_parameter(SearchParameter {
        name: "name".to_string(),
        param_type: SearchParamType::String,
        modifier: Some(SearchModifier::Contains),
        values: vec![SearchValue::eq("smith")],
        chain: vec![],
    })
    .with_sort(SortDirective::parse("-_lastUpdated"))
    .with_count(20);

// With _include
let query_with_include = SearchQuery::new("Observation")
    .with_include(IncludeDirective {
        include_type: IncludeType::Include,
        source_type: "Observation".to_string(),
        search_param: "patient".to_string(),
        target_type: Some("Patient".to_string()),
        iterate: false,
    });

Backend Capability Matrix

The matrix below shows which FHIR operations each backend supports. This reflects the actual implementation status, not aspirational goals.

For a capability-by-capability narrative of FHIR Search against the spec — including the REST-layer vs. backend boundary and a roadmap of known gaps — see docs/search-spec-assessment.md.

Note: Documentation links reference build.fhir.org, which contains the current FHIR development version. Some features marked as planned are new and may be labeled "Trial Use" in the specification.

Legend: ✓ Implemented | ◐ Partial | ○ Planned | ✗ Not planned | † Requires external service

Feature	SQLite	PostgreSQL	MongoDB	Cassandra	Neo4j	Elasticsearch	S3
Core Operations
CRUD	✓	✓	✓	○	○	✓	✓
Versioning (vread)	✓	✓	✓	○	○	○	✓
Optimistic Locking	✓	✓	✓	○	○	✗	✓
Instance History	✓	✓	○	✗	○	✗	✓
Type History	✓	✓	○	✗	○	✗	✓
System History	✓	✓	○	✗	○	✗	✓
Batch Bundles	✓	✓	✓	○	○	○	✓
Transaction Bundles	✓	✓	✓	✗	○	✗	◐
Conditional Operations	✓	✓	✓	✗	○	○	✗
Conditional Patch	✓	✓	○	✗	○	○	✗
Delete History	✓	✓	○	✗	○	✗	✗
Multitenancy
Shared Schema	✓	✓	✓	○	○	✓	✓
Schema-per-Tenant	✗	○	○	✗	✗	✗	✗
Database-per-Tenant	✓	○	○	○	○	○	✓
Row-Level Security	✗	○	✗	✗	✗	✗	✗
Search Parameters
String	✓	✓	✓	✗	○	✓	✗
Token	✓	✓	✓	○	○	✓	✗
Reference	✓	✓	✓	✗	○	✓	✗
Date	✓	✓	✓	○	○	✓	○
Number	✓	✓	✓	✗	○	✓	○
Quantity	✓	✓	✓	✗	✗	✓	○
URI	✓	✓	✓	○	○	✓	○
Composite	✓	✓	○	✗	○	✓	✗
Search Modifiers
:exact	✓	✓	✓	○	○	✓	○
:contains	✓	✓	✓	✗	○	✓	✗
:text (full-text)	✓	◐	○	✗	✗	✓	✗
:not	✓	✓	○	✗	○	✓	○
:missing	✓	✓	○	✗	○	✓	○
:above / :below	◐	◐	◐	✗	○	◐	✗
:in / :not-in	✗	†○	†○	✗	○	†○	✗
:of-type	✓	✓	○	✗	○	✓	✗
:text-advanced	✓	†○	†○	✗	✗	✓	✗
Special Parameters
_text (narrative search)	✓	✓	○	✗	✗	✓	✗
_content (full content)	✓	✓	○	✗	✗	✓	✗
_filter (advanced filtering)	✓	○	○	✗	○	○	✗
Advanced Search
Chained Parameters	✓	✓	◐	✗	○	◐	✗
Reverse Chaining (_has)	✓	✓	◐	✗	○	◐	✗
_include	✓	✓	✓	✗	○	✓	✗
_revinclude	✓	✓	✓	✗	○	✓	✗
Pagination
Offset	✓	✓	✓	✗	○	✓	✗
Cursor (keyset)	✓	✓	✓	○	○	✓	○
Sorting
Single field	✓	✓	✓	✗	○	✓	✗
Multiple fields	✓	✓	◐	✗	○	✓	✗
Bulk Operations
Bulk Export	✓	✓	○	○	○	○	◐
Bulk Submit ingest	✓	✓	○	○	○	○	✓
Bulk Submit REST worker	✓	✓	✗	✗	✗	✗	✗

Notes on partial cells:

• Sorting — SQLite and PostgreSQL sort by any indexed search parameter (string, token, date, number, quantity, reference, URI) via a correlated subquery into the search index, taking the min value ascending / max descending for multi-valued params; _id/_lastUpdated sort on the resources table directly. Cursor (keyset) pagination is consistent with the active sort: the sort key value is encoded into the opaque cursor and the keyset comparison runs on it, so deep paging preserves the sort order. A multi-field _sort returns a single page (no cursor). MongoDB sorts by _id/_lastUpdated only and cannot combine a custom sort with cursor pagination, hence ◐ for multiple fields.
• :above / :below — two mechanisms (◐ = both, conditional on context): (1) hierarchical URI prefix matching is native to SQLite, PostgreSQL, and Elasticsearch (no external service); (2) token/code hierarchy (e.g. code:below=http://snomed.info/sct|73211009) is resolved at the REST layer — the code and its descendants (:below, is-a) or ancestors (:above, generalizes) are expanded via the terminology server's $expand, then matched as a plain token OR list on any backend. Token hierarchy therefore requires HFS_TERMINOLOGY_SERVER.
• :in / :not-in — handled at the REST layer: :in is expanded against a terminology server before the query reaches the backend; :not-in returns 501 Not Implemented. No backend resolves these natively.
• Chained / _has — SQLite and PostgreSQL resolve chains natively in-backend (✓). For every other backend (◐), the REST layer resolves chained and reverse-chained parameters via application-side joins (search::resolve_chains): each hop is one plain search() against the backend, and the result is folded into an _id filter. So GET /Observation?subject.name=Smith and ?_has:Observation:subject:code=… work end-to-end on every searchable backend, including Elasticsearch and MongoDB.
• Composite — SQLite, PostgreSQL, and Elasticsearch evaluate composite component values (token, string, number, quantity, date) end-to-end (✓): the REST layer resolves component types from the registry and the extractor indexes each composite instance as a composite_group. SQLite/PG match all components within one group via GROUP BY … HAVING; Elasticsearch indexes each instance as one nested object with inline component values and matches with a single nested query. See docs/search-spec-assessment.md.

The S3 backend is intentionally storage-focused (CRUD/version/history and BulkSubmitProvider ingestion) and does not act as a full FHIR search engine. For bulk export, S3 can feed system-level batches through ExportDataProvider and can store output files through S3OutputStore, but job state belongs to SQLite or PostgreSQL. $bulk-submit REST worker/job state also belongs to SQLite or PostgreSQL; S3 supports only the synchronous ingest provider. Patient-level and Group-level export compartment enumeration are not supported by S3 as the resource store. For query-heavy deployments, use a DB/search backend as primary query engine and compose S3 as archive/history/output storage.

Primary/Secondary Role Matrix

Backends can serve as primary (CRUD, versioning, transactions) or secondary (optimized for specific query patterns). When a secondary search backend is configured, the primary backend's search indexing is automatically disabled to avoid data duplication.

Configuration	Primary	Secondary	Status	Use Case
SQLite alone	SQLite	—	✓ Implemented	Development, testing, small deployments
SQLite + Elasticsearch	SQLite	Elasticsearch (search)	✓ Implemented	Small prod with robust search
PostgreSQL alone	PostgreSQL	—	✓ Implemented	Production OLTP
PostgreSQL + Elasticsearch	PostgreSQL	Elasticsearch (search)	✓ Implemented	OLTP + advanced search
PostgreSQL + Neo4j	PostgreSQL	Neo4j (graph)	Planned	Graph-heavy queries
Cassandra alone	Cassandra	—	Planned	High write throughput
Cassandra + Elasticsearch	Cassandra	Elasticsearch (search)	Planned	Write-heavy + search
MongoDB alone	MongoDB	—	✓ Implemented	Document-centric
MongoDB + Elasticsearch	MongoDB	Elasticsearch (search)	✓ Implemented	Document-centric + offloaded search
S3 alone	S3	—	✓ Implemented (storage-focused)	Archival/history storage
S3 + Elasticsearch	S3	Elasticsearch (search)	✓ Implemented	Large-scale + search

Backend Selection Guide

Use Case	Recommended Backend	Rationale
Development & Testing	SQLite	Zero configuration, in-memory mode
Production OLTP	PostgreSQL	ACID transactions, JSONB, mature ecosystem
Document-centric	MongoDB	Natural FHIR alignment, flexible schema
Graph queries	Neo4j	Efficient relationship traversal
Full-text search	Elasticsearch	Optimized inverted indexes, analyzers
Bulk analytics	S3 + Parquet	Cost-effective, columnar, ML-ready
High write throughput	Cassandra	Distributed writes, eventual consistency

Feature Flags

Feature	Description	Driver
sqlite (default)	SQLite (in-memory and file)	rusqlite
postgres	PostgreSQL with JSONB	tokio-postgres
cassandra	Apache Cassandra	cdrs-tokio
mongodb	MongoDB document store	mongodb
neo4j	Neo4j graph database	neo4rs
elasticsearch	Elasticsearch search	elasticsearch
s3	AWS S3 object storage	aws-sdk-s3

Building & Running Storage Backends

This section covers building the hfs binary with specific backend support and setting up the required infrastructure.

SQLite (Default)

Zero-configuration setup — no external dependencies required.

# Build with default SQLite backend
cargo build --bin hfs --release

# Run
./target/release/hfs

SQLite handles all CRUD operations, versioning, history, and search using its built-in FTS5 full-text search engine. Data is stored in fhir.db by default.

SQLite + Elasticsearch

SQLite handles CRUD, versioning, history, and transactions. Elasticsearch handles all search operations with:

• Full-text search with relevance scoring (_text, _content)
• All FHIR search parameter types (string, token, date, number, quantity, reference, URI, composite)
• Advanced text search with stemming, boolean operators, and proximity matching (:text-advanced)
• Cursor-based pagination via search_after

Prerequisites: A running Elasticsearch 8.x instance.

# Build with Elasticsearch support
cargo build --bin hfs --features sqlite,elasticsearch --release

# Start Elasticsearch (example using Docker)
docker run -d --name es -p 9200:9200 \
  -e "discovery.type=single-node" \
  -e "xpack.security.enabled=false" \
  elasticsearch:8.15.0

# Start the server
HFS_STORAGE_BACKEND=sqlite-elasticsearch \
HFS_ELASTICSEARCH_NODES=http://localhost:9200 \
  ./target/release/hfs

PostgreSQL

Full-featured relational backend for production deployments with JSONB storage, full-text search, and advanced multi-tenant isolation strategies.

• Full CRUD operations with ACID transactions
• Full-text search via PostgreSQL's tsvector/tsquery
• All FHIR search parameter types (string, token, date, number, quantity, reference, URI, composite)
• Chained parameters and reverse chaining (_has)
• _include and _revinclude resolution
• Multi-tenant support (shared schema, schema-per-tenant, database-per-tenant)

Prerequisites: A running PostgreSQL instance (14+).

# Build with PostgreSQL support
cargo build --bin hfs --features postgres --release

# Start PostgreSQL (example using Docker)
docker run -d --name pg -p 5432:5432 \
  -e POSTGRES_USER=hfs \
  -e POSTGRES_PASSWORD=hfs \
  -e POSTGRES_DB=fhir \
  postgres:16

# Start the server
HFS_STORAGE_BACKEND=postgres \
HFS_DATABASE_URL="postgresql://hfs:hfs@localhost:5432/fhir" \
  ./target/release/hfs

PostgreSQL + Elasticsearch

PostgreSQL handles CRUD, versioning, history, and transactions with ACID guarantees. Elasticsearch handles all search operations. Combines PostgreSQL's production-grade storage with Elasticsearch's search capabilities.

• Full CRUD operations with ACID transactions via PostgreSQL
• Full-text search with relevance scoring (_text, _content) via Elasticsearch
• All FHIR search parameter types (string, token, date, number, quantity, reference, URI, composite)
• Advanced text search with stemming, boolean operators, and proximity matching (:text-advanced)
• Multi-tenant support (shared schema, schema-per-tenant, database-per-tenant)

Prerequisites: Running PostgreSQL (14+) and Elasticsearch 8.x instances.

# Build with PostgreSQL and Elasticsearch support
cargo build --bin hfs --features postgres,elasticsearch --release

# Start PostgreSQL (example using Docker)
docker run -d --name pg -p 5432:5432 \
  -e POSTGRES_USER=hfs \
  -e POSTGRES_PASSWORD=hfs \
  -e POSTGRES_DB=fhir \
  postgres:16

# Start Elasticsearch (example using Docker)
docker run -d --name es -p 9200:9200 \
  -e "discovery.type=single-node" \
  -e "xpack.security.enabled=false" \
  elasticsearch:8.15.0

# Start the server
HFS_STORAGE_BACKEND=postgres-elasticsearch \
HFS_DATABASE_URL="postgresql://hfs:hfs@localhost:5432/fhir" \
HFS_ELASTICSEARCH_NODES=http://localhost:9200 \
  ./target/release/hfs

MongoDB

MongoDB provides document-centric primary storage with full FHIR capabilities including CRUD, versioning, history, search, and transactions.

• Full CRUD operations with document-native resource storage
• Versioning and history providers (vread, instance/type/system history)
• Transaction bundles with urn:uuid reference resolution (requires replica set)
• Native search (string, token, reference, date, number, quantity, URI parameters; composite parameters, _text/_content, and most modifiers beyond :exact/:contains are not yet supported; chained/_has work via the REST-layer resolver)
• _include and _revinclude resolution
• Conditional create, update, and delete operations
• Cursor and offset pagination; sorting by _id/_lastUpdated (a custom sort cannot be combined with cursor pagination)
• Shared-schema multitenancy with strict tenant filtering
• Optimistic locking with ETag support

Prerequisites: A running MongoDB instance. Use standalone for basic deployments or replica set/sharded topology for transaction bundle support.

# Build with MongoDB support
cargo build --bin hfs --features mongodb --release

# Start MongoDB (example using Docker)
docker run -d --name mongo -p 27017:27017 \
  mongo:8.0

# Start the server
HFS_STORAGE_BACKEND=mongodb \
HFS_DATABASE_URL="mongodb://localhost:27017" \
HFS_MONGODB_DATABASE=helios \
  ./target/release/hfs

MongoDB runtime configuration also supports:

• HFS_MONGODB_URL or HFS_MONGODB_URI as preferred connection-string inputs
• HFS_MONGODB_DATABASE to select the database name (default: helios)
• HFS_MONGODB_MAX_CONNECTIONS to control the driver pool size (default: 10)
• HFS_MONGODB_CONNECT_TIMEOUT_MS to control the connection timeout (default: 5000)

MongoDB + Elasticsearch

MongoDB remains the canonical write/read store while Elasticsearch owns delegated search execution. This mode mirrors the existing SQLite + Elasticsearch and PostgreSQL + Elasticsearch composite patterns.

• MongoDB handles CRUD, versioning, history, and conditional write behavior
• Elasticsearch handles delegated search queries, including full-text search
• MongoDB search index population is automatically disabled via search_offloaded
• Composite routing preserves MongoDB as the source of truth for reads and writes

Prerequisites: Running MongoDB and Elasticsearch 8.x instances.

# Build with MongoDB and Elasticsearch support
cargo build --bin hfs --features mongodb,elasticsearch --release

# Start MongoDB (example using Docker)
docker run -d --name mongo -p 27017:27017 \
  mongo:8.0

# Start Elasticsearch (example using Docker)
docker run -d --name es -p 9200:9200 \
  -e "discovery.type=single-node" \
  -e "xpack.security.enabled=false" \
  elasticsearch:8.15.0

# Start the server
HFS_STORAGE_BACKEND=mongodb-elasticsearch \
HFS_DATABASE_URL="mongodb://localhost:27017" \
HFS_MONGODB_DATABASE=helios \
HFS_ELASTICSEARCH_NODES=http://localhost:9200 \
  ./target/release/hfs

S3 + Elasticsearch

S3 handles CRUD, versioning, history, and bulk-submit artifacts. Elasticsearch handles all search operations. For bulk export, this topology can use S3 as the resource data provider for system-level exports and S3OutputStore as the output-file store; export job state still lives in the configured SQLite or PostgreSQL bulk-export job store.

• CRUD persistence via S3 objects (current pointer + immutable history versions)
• Versioning (vread, optimistic locking via version checks)
• Instance, type, and system history via immutable history objects
• Batch bundles and best-effort transaction bundles
• Bulk export data provider for system-level exports
• Optional S3 bulk-export output files via S3OutputStore
• Bulk submit with rollback change log
• Full-text search with relevance scoring (_text, _content) via Elasticsearch
• All FHIR search parameter types (string, token, date, number, quantity, reference, URI, composite)
• Advanced text search with stemming, boolean operators, and proximity matching (:text-advanced)
• Tenant isolation (PrefixPerTenant or BucketPerTenant)

Prerequisites: An AWS S3 bucket (or S3-compatible service) and a running Elasticsearch 8.x instance.

# Build with S3 and Elasticsearch support
cargo build --bin hfs --features s3,elasticsearch --release

# Start Elasticsearch (example using Docker)
docker run -d --name es -p 9200:9200 \
  -e "discovery.type=single-node" \
  -e "xpack.security.enabled=false" \
  elasticsearch:8.15.0

# Start the server (AWS S3)
HFS_STORAGE_BACKEND=s3-elasticsearch \
HFS_S3_BUCKET=my-fhir-bucket \
HFS_ELASTICSEARCH_NODES=http://localhost:9200 \
  ./target/release/hfs

S3 Environment Variables

Variable	Default	Description
HFS_S3_BUCKET	hfs	S3 bucket name
HFS_S3_REGION	(provider chain)	AWS region override
HFS_S3_PREFIX	(none)	Optional global key prefix
HFS_S3_VALIDATE_BUCKETS	true	Validate buckets on startup via HeadBucket

AWS credentials are resolved via the standard AWS provider chain (AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY, EC2 instance metadata, shared credentials file, SSO, etc.).

S3-Compatible Endpoints (MinIO, etc.)

For S3-compatible services, configure the endpoint programmatically via S3BackendConfig:

use helios_persistence::backends::s3::{S3BackendConfig, S3TenancyMode};

let config = S3BackendConfig {
    tenancy_mode: S3TenancyMode::PrefixPerTenant {
        bucket: "minio-bucket".to_string(),
    },
    endpoint_url: Some("http://127.0.0.1:9000".to_string()),
    allow_http: true,
    force_path_style: true,
    ..Default::default()
};

When endpoint_url is set, the backend automatically defaults force_path_style to true and region to us-east-1 if not otherwise specified.

Key Differences from SQLite/PG + ES

Unlike SQLite and PostgreSQL, the S3 backend has no built-in search parameter registry. When composing S3 + Elasticsearch, the ES backend creates its own standalone registry with minimal embedded search parameters (_id, _lastUpdated, _tag, _profile, _security). For full search capability, use with_shared_registry() with parameters loaded from spec files.

use std::collections::HashMap;
use std::sync::Arc;
use helios_persistence::backends::elasticsearch::{ElasticsearchBackend, ElasticsearchConfig};
use helios_persistence::backends::s3::{S3Backend, S3BackendConfig};
use helios_persistence::composite::{CompositeConfig, CompositeStorage, DynStorage, DynSearchProvider};
use helios_persistence::core::BackendKind;

// Create S3 backend
let s3_config = S3BackendConfig::default();
let s3 = Arc::new(S3Backend::from_env_async(s3_config).await?);

// Create ES backend (standalone registry — S3 has no registry to share)
let es_config = ElasticsearchConfig::default();
let es = Arc::new(ElasticsearchBackend::new(es_config)?);

// Build composite
let composite_config = CompositeConfig::builder()
    .primary("s3", BackendKind::S3)
    .search_backend("es", BackendKind::Elasticsearch)
    .build()?;

let mut backends = HashMap::new();
backends.insert("s3".to_string(), s3.clone() as DynStorage);
backends.insert("es".to_string(), es.clone() as DynStorage);

let mut search_providers = HashMap::new();
search_providers.insert("s3".to_string(), s3.clone() as DynSearchProvider);
search_providers.insert("es".to_string(), es.clone() as DynSearchProvider);

let composite = CompositeStorage::new(composite_config, backends)?
    .with_search_providers(search_providers)
    .with_full_primary(s3);

How Search Offloading Works

When HFS_STORAGE_BACKEND is set to sqlite-elasticsearch, postgres-elasticsearch, mongodb-elasticsearch, or s3-elasticsearch, the server:

1. Creates the primary backend (SQLite, PostgreSQL, MongoDB, or S3). For SQLite/PG/MongoDB, search indexing is disabled; S3 has no search indexing to disable.
2. Creates an Elasticsearch backend. For SQLite/PG/MongoDB, it shares the primary backend's search parameter registry; for S3, it creates its own standalone registry.
3. Wraps both in a CompositeStorage that routes:
   • All writes (create, update, delete, conditional ops, transactions) → primary backend, then syncs to ES
   • All reads (read, vread, history) → primary backend
   • All search operations → Elasticsearch

This is controlled by the search_offloaded flag on the primary backend, which the composite layer sets automatically when a search secondary is configured.

Composite Usage

use std::collections::HashMap;
use std::sync::Arc;
use helios_persistence::composite::{CompositeConfig, CompositeStorage, DynStorage, DynSearchProvider};
use helios_persistence::core::BackendKind;
use helios_persistence::backends::sqlite::SqliteBackend;
use helios_persistence::backends::elasticsearch::{ElasticsearchBackend, ElasticsearchConfig};

// Create backends
let mut sqlite = SqliteBackend::new("fhir.db")?;
sqlite.set_search_offloaded(true);  // Disable SQLite search indexing
let sqlite = Arc::new(sqlite);

let es = Arc::new(ElasticsearchBackend::with_shared_registry(
    ElasticsearchConfig::default(),
    sqlite.search_registry().clone(),
)?);

// Build composite
let config = CompositeConfig::builder()
    .primary("sqlite", BackendKind::Sqlite)
    .search_backend("es", BackendKind::Elasticsearch)
    .build()?;

let mut backends = HashMap::new();
backends.insert("sqlite".to_string(), sqlite.clone() as DynStorage);
backends.insert("es".to_string(), es.clone() as DynStorage);

let mut search_providers = HashMap::new();
search_providers.insert("sqlite".to_string(), sqlite.clone() as DynSearchProvider);
search_providers.insert("es".to_string(), es.clone() as DynSearchProvider);

// with_full_primary() enables delegation of ConditionalStorage, VersionedStorage,
// InstanceHistoryProvider, and BundleProvider through the composite layer.
let composite = CompositeStorage::new(config, backends)?
    .with_search_providers(search_providers)
    .with_full_primary(sqlite);

S3 Backend

The S3 backend is a storage-focused persistence backend using AWS S3 object storage. It handles CRUD, versioning/history, and synchronous bulk-submit ingest provider workflows, but is intentionally not a FHIR search engine. For bulk export, S3 participates in two narrower roles: S3Backend can provide resource batches for system-level exports, and S3OutputStore can store finalized NDJSON output files. Bulk-export and $bulk-submit REST worker job state, progress, manifests, leases, and file metadata are not stored in S3; they live in SQLite or PostgreSQL.

Scope

Primary responsibilities:

• CRUD persistence of resources
• Versioning (vread, list_versions, optimistic conflict checks)
• Instance/type/system history via immutable history objects plus history index events
• Batch bundles and best-effort transaction bundles (non-atomic with compensating rollback)
• Bulk export resource data provider for system-level exports
• Bulk export output storage through S3OutputStore when configured separately from job state
• Bulk submit (ingest + raw artifact persistence + rollback change log)
• Tenant isolation (PrefixPerTenant or BucketPerTenant)

Explicit non-goals: Advanced FHIR search semantics (date/number/quantity comparisons, chained query planning, _has, include/revinclude fanout, cursor keyset queries).

Configuration

use helios_persistence::backends::s3::{S3BackendConfig, S3TenancyMode};

let config = S3BackendConfig {
    tenancy_mode: S3TenancyMode::PrefixPerTenant {
        bucket: "hfs".to_string(),
    },
    prefix: None,
    region: None,
    validate_buckets_on_startup: true,
    bulk_submit_batch_size: 100,
    ..Default::default()
};

Option	Default	Description
tenancy_mode	PrefixPerTenant { bucket: "hfs" }	Tenant-to-bucket mapping strategy
prefix	None	Optional global key prefix applied before backend keys
region	None	AWS region override (falls back to provider chain)
validate_buckets_on_startup	true	Validate configured buckets with HeadBucket on startup
bulk_submit_batch_size	100	Default ingestion batch size for bulk submit processing

Tenancy Modes

Mode	Description
PrefixPerTenant	All tenants share one bucket with tenant-specific key prefixes
BucketPerTenant	Each tenant maps to a specific bucket via an explicit tenant→bucket map

Object Model

Resource objects:

Object	Key Pattern
Current pointer	.../resources/{type}/{id}/current.json
Immutable history version	.../resources/{type}/{id}/_history/{version}.json
Type history event	.../history/type/{type}/{ts}_{id}_{version}_{suffix}.json
System history event	.../history/system/{ts}_{type}_{id}_{version}_{suffix}.json

Bulk export output objects:

Object	Key Pattern
Finalized NDJSON part	{tenant_id}/exports/{job_id}/{file_type}-{resource_type}-{part_index}-{fencing_token}.ndjson

Bulk-export job state is deliberately not an S3 object model. SQLite and PostgreSQL store the job row, progress, leases/fencing tokens, file metadata, and raw manifest rows. S3OutputStore stores only finalized output parts and deletes every object under {tenant_id}/exports/{job_id}/ during cancellation or retention cleanup. The REST layer assembles the client-facing manifest from the job store plus ExportOutputStore::download_url.

Bulk submit objects:

Object	Key Pattern
Submission state	.../bulk/submit/{submitter}/{submission_id}/state.json
Manifest	.../bulk/submit/{submitter}/{submission_id}/manifests/{manifest_id}.json
Raw input	.../bulk/submit/{submitter}/{submission_id}/raw/{manifest_id}/line-{line}.ndjson
Results	.../bulk/submit/{submitter}/{submission_id}/results/{manifest_id}/line-{line}.json
Change log	.../bulk/submit/{submitter}/{submission_id}/changes/{change_id}.json

Consistency and Transaction Notes

• The backend never creates buckets — startup/runtime bucket checks use HeadBucket only.
• Optimistic locking relies on version checks plus S3 preconditions (If-Match, If-None-Match) where applicable.
• Transaction bundle behavior is best-effort: entries are applied sequentially, rollback is attempted in reverse order on failure, but rollback is not guaranteed under concurrent writes or partial failures.

AWS Credentials and Region

Uses the AWS SDK for Rust (aws_sdk_s3) with standard provider chain:

• Region may be provided in config or via AWS_REGION
• Environment credentials (AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, optional AWS_SESSION_TOKEN) are supported by provider chain behavior

Endpoint Modes

AWS S3 mode

Use this mode when connecting to AWS-managed S3 endpoints.

• endpoint_url = None
• allow_http = false (default)
• force_path_style = false (default)

use helios_persistence::backends::s3::{S3BackendConfig, S3TenancyMode};

let config = S3BackendConfig {
    tenancy_mode: S3TenancyMode::PrefixPerTenant {
        bucket: "my-aws-bucket".to_string(),
    },
    endpoint_url: None,
    force_path_style: false,
    allow_http: false,
    ..Default::default()
};

S3-compatible endpoint mode (MinIO, etc.)

Use this mode for custom endpoints.

• set endpoint_url
• set allow_http=true for local http://... endpoints
• path-style is defaulted on when endpoint mode is active
• region falls back to us-east-1 when not provided

use helios_persistence::backends::s3::{S3BackendConfig, S3TenancyMode};

let config = S3BackendConfig {
    tenancy_mode: S3TenancyMode::PrefixPerTenant {
        bucket: "minio-bucket".to_string(),
    },
    endpoint_url: Some("http://127.0.0.1:9000".to_string()),
    allow_http: true,
    force_path_style: true,
    ..Default::default()
};

Notes:

• http:// endpoints are rejected unless allow_http=true.
• AWS behavior is unchanged when endpoint_url is not set.
• Buckets are never created by production backend code; startup validation uses HeadBucket.

MinIO Integration Tests

MinIO parity tests live in:

• crates/persistence/tests/minio_s3_tests.rs

The suite is opt-in and env-gated:

• RUN_MINIO_S3_TESTS=1

Optional overrides:

• MINIO_IMAGE (default: minio/minio)
• MINIO_TAG (default: RELEASE.2025-02-28T09-55-16Z)
• MINIO_ROOT_USER (default: minioadmin)
• MINIO_ROOT_PASSWORD (default: minioadmin)
• HFS_MINIO_TEST_BUCKET (if unset, tests auto-generate a unique bucket)

Example:

MinIO Testing

RUN_MINIO_S3_TESTS=1 \
cargo test -p helios-persistence --features s3 --test minio_s3_tests

S3 Testing

Note: Make sure aws sso login is set up and running before executing S3 Testing

export RUN_AWS_S3_TESTS=1
export HFS_S3_TEST_BUCKET="your-existing-bucket-name"
export AWS_REGION="us-east-1"   # or your bucket’s region
cargo test -p helios-persistence --test s3_tests --features s3

Implementation Status

Phase 1: Core Types ✓

• Error types with comprehensive variants
• Tenant types (TenantId, TenantContext, TenantPermissions)
• Stored resource types with versioning metadata
• Search parameter types (all FHIR parameter types)
• Pagination types (cursor and offset)

Phase 2: Core Traits ✓

• Backend trait with capability discovery
• ResourceStorage trait (CRUD operations)
• VersionedStorage trait (vread, If-Match)
• History provider traits (instance, type, system)
• Search provider traits (basic, chained, _include, terminology)
• Transaction traits (ACID, bundles)
• Capabilities trait (CapabilityStatement generation)

Phase 3: Tenancy Strategies ✓

• Shared schema strategy with RLS support
• Schema-per-tenant strategy with PostgreSQL search_path
• Database-per-tenant strategy with pool management

Phase 4: SQLite Backend ✓

• Connection pooling (r2d2)
• Schema migrations
• ResourceStorage implementation
• VersionedStorage implementation
• History providers (instance, type, system)
• TransactionProvider implementation
• Conditional operations (conditional create/update/delete)

Transaction & Batch Support ◐

FHIR transaction and batch bundle processing.

Backend Support: Transaction bundles require ACID support. SQLite supports transactions. Cassandra, Elasticsearch, and S3 do not support transactions (batch only). See the capability matrix above.

Implemented Features:

• Transaction bundles - Atomic all-or-nothing processing with automatic rollback on failure
• Batch bundles - Independent entry processing (failures don't affect other entries)
• Processing order - Entries processed per FHIR spec: DELETE → POST → PUT/PATCH → GET
• Reference resolution - urn:uuid: references automatically resolved to assigned IDs after creates
• fullUrl support - Track temporary identifiers for intra-bundle references
• Conditional headers - If-Match, If-None-Match, If-None-Exist in bundle entries
• Error responses - Transaction failures return OperationOutcome with failing entry index
• Response ordering - Results returned in original request entry order

Not Yet Implemented:

Gap	Description	Spec Reference
Conditional reference resolution	References like Patient?identifier=12345 should resolve via search	Transaction
PATCH method	PATCH operations in bundle entries return 501	Patch
Duplicate resource detection	Same resource appearing twice in transaction should fail	Transaction
Prefer header handling	return=minimal, return=representation, return=OperationOutcome	Prefer
History bundle acceptance	Servers SHOULD accept history bundles for replay	History
Version-specific references	resolve-as-version-specific extension support	References
lastModified in response	Bundle entry responses should include lastModified	Transaction

SQLite Search Implementation ✓

The SQLite backend includes a complete FHIR search implementation using pre-computed indexes:

Search Parameter Registry & Extraction:

• SearchParameterRegistry - In-memory cache of active SearchParameter definitions
• SearchParameterLoader - Loads embedded R4 standard parameters at startup
• SearchParameterExtractor - FHIRPath-based value extraction using helios-fhirpath
• Dynamic SearchParameter handling - POST/PUT/DELETE to SearchParameter updates the registry

Search Index & Query:

• Pre-computed search_index table for fast queries
• All 8 parameter type handlers (string, token, date, number, quantity, reference, URI, composite)
• Modifier support (:exact, :contains, :missing, :not, :identifier, :below, :above)
• Prefix support for date/number/quantity (eq, ne, gt, lt, ge, le, sa, eb, ap)
• _include and _revinclude resolution
• Cursor-based and offset pagination
• Single-field sorting

Full-Text Search (FTS5):

• resource_fts FTS5 virtual table for full-text indexing
• Narrative text extraction from text.div with HTML stripping
• Full content extraction from all resource string values
• _text parameter - searches narrative content
• _content parameter - searches all resource text
• :text-advanced modifier - advanced FTS5-based search with:
  • Porter stemming (e.g., "run" matches "running")
  • Boolean operators (AND, OR, NOT)
  • Phrase matching ("heart failure")
  • Prefix search (cardio*)
  • Proximity matching (NEAR operator)
• Porter stemmer tokenization for improved search quality
• Automatic FTS indexing on resource create/update/delete

Chained Parameters & Reverse Chaining:

• N-level forward chains (e.g., Observation?subject.organization.name=Hospital)
• Nested reverse chains / _has (e.g., Patient?_has:Observation:subject:code=1234-5)
• Type modifiers for ambiguous references (e.g., subject:Patient.name=Smith)
• SQL-based chain resolution using efficient nested subqueries
• Registry-based type inference with fallback heuristics
• Configurable depth limits (default: 4, max: 8)

Reindexing:

• ReindexableStorage trait for backend-agnostic reindexing
• ReindexOperation with background task execution
• Progress tracking and cancellation support
• $reindex HTTP endpoint (planned for server layer)

Capability Reporting:

• SearchCapabilityProvider implementation
• Runtime capability discovery from registry

Bulk Operations:

• BulkExportStorage trait implementation (FHIR Bulk Data Access IG)
  • System-level export (/$export)
  • Patient-level export (/Patient/$export)
  • Group-level export (/Group/[id]/$export)
  • Job lifecycle management (pending, in-progress, completed, failed, cancelled)
  • Streaming NDJSON batch generation
  • Type filtering and _since parameter support
• BulkSubmitProvider trait implementation (FHIR Bulk Submit)
  • Submission lifecycle management
  • Manifest creation and management
  • Entry processing with validation
  • Rollback support for failed submissions
• Schema migration v5 to v6 with 7 new tables for bulk operations

Phase 5: Elasticsearch Backend ✓

• Backend structure with connection management and health checks
• Index schema and mappings (nested objects for multi-value search params)
• ResourceStorage implementation for composite sync support
• Search query translation (FHIR SearchQuery → ES Query DSL)
• Parameter type handlers (string, token, date, number, quantity, reference, URI, composite — each composite instance is one nested object with inline component values, matched by a single nested query)
• Full-text search (_text, _content, :text-advanced)
• Modifier support (:exact, :contains, :text, :not, :missing, :above, :below, :of-type)
• _include and _revinclude resolution
• Cursor-based (search_after) and offset pagination
• Multi-field sorting
• Search offloading: when Elasticsearch is the search secondary, the primary backend skips search index population

Phase 5b: PostgreSQL Backend ✓

• Connection pooling (deadpool-postgres)
• Schema migrations with JSONB storage
• ResourceStorage implementation (CRUD)
• VersionedStorage implementation (vread, If-Match)
• History providers (instance, type, system)
• TransactionProvider with configurable isolation levels
• Conditional operations (conditional create/update/delete)
• SearchProvider with all parameter types including composite (string, token, reference, date, number, quantity, URI, composite; supports :exact/:contains/:not/:missing/:of-type and URI :above/:below; the :text-advanced modifier is not yet implemented). Composite search works end-to-end (REST → registry-resolved components → grouped index → query).
• ChainedSearchProvider and reverse chaining (_has)
• Full-text search (tsvector/tsquery)
• _sort by _id/_lastUpdated and any indexed search parameter via a search_index correlated subquery (first-page and offset paths)
• _include and _revinclude resolution
• BulkExportStorage and BulkSubmitProvider
• Search offloading support
• ReindexableStorage implementation

Phase 5c: S3 Backend ✓

• S3BackendConfig with PrefixPerTenant and BucketPerTenant tenancy modes
• ResourceStorage implementation (CRUD via S3 objects)
• VersionedStorage implementation (vread, optimistic locking)
• History providers (instance, type, system via immutable history objects)
• Batch and best-effort transaction bundles
• ExportDataProvider implementation for system-level bulk export
• S3OutputStore implementation for bulk-export NDJSON output files
• BulkSubmitProvider implementation (ingest, raw artifacts, rollback change log)

Phase 5+: Additional Backends (Planned)

• Cassandra backend (wide-column, partition keys)
• MongoDB Phase 1 scaffold (module wiring, config, Backend trait baseline)
• MongoDB Phase 2 core storage parity (CRUD/count/read_batch/create_or_update, tenant isolation, soft-delete, schema bootstrap)
• MongoDB Phase 3 versioning/history plus best-effort session-backed consistency
• MongoDB Phase 4 native search, pagination/sorting, and conditional create/update/delete
• MongoDB Phase 5 composite MongoDB + Elasticsearch integration and runtime wiring
• MongoDB Phase 6 runtime wiring verification, documentation sync, and release-readiness validation
• Neo4j backend (graph queries, Cypher)

Phase 6: Composite Storage ✓

• Query analysis and feature detection
• Multi-backend coordination with primary-secondary model
• Cost-based query routing
• Result merging strategies
• Secondary backend synchronization
• Health monitoring
• Configuration Advisor HTTP API
• Full primary delegation via with_full_primary() — CompositeStorage now implements ConditionalStorage, VersionedStorage, InstanceHistoryProvider, and BundleProvider by delegating to the primary backend

Composite Storage

The composite storage layer enables polyglot persistence by coordinating multiple database backends for optimal FHIR resource storage and querying.

Design Principles

1. Single Source of Truth: One primary backend handles all FHIR resource CRUD operations, versioning, and history. This is the authoritative store.

2. Feature-Based Routing: Queries are automatically routed based on detected features (chained search, full-text, terminology) to appropriate backends.

3. Eventual Consistency: Secondary backends may lag behind primary (configurable sync/async modes with documented consistency guarantees).

4. Graceful Degradation: If a secondary backend is unavailable, the system falls back to primary with potentially degraded performance.

Valid Backend Configurations

Configuration	Primary	Secondary(s)	Status	Use Case
SQLite-only	SQLite	None	✓ Implemented	Development, testing, small deployments
SQLite + ES	SQLite	Elasticsearch	✓ Implemented	Small prod with robust search
PostgreSQL-only	PostgreSQL	None	✓ Implemented	Production OLTP
PostgreSQL + ES	PostgreSQL	Elasticsearch	✓ Implemented	OLTP + advanced search
PostgreSQL + Neo4j	PostgreSQL	Neo4j	Planned	Graph-heavy queries
MongoDB-only	MongoDB	None	✓ Implemented	Document-centric primary
MongoDB + ES	MongoDB	Elasticsearch	✓ Implemented	Document-centric + search
S3 alone	S3	—	✓ Implemented	Archival/history storage
S3 + ES	S3	Elasticsearch	✓ Implemented	Large-scale + search

Quick Start

use helios_persistence::composite::{
    CompositeConfigBuilder, BackendRole, SyncMode,
};
use helios_persistence::core::BackendKind;

// Development configuration (SQLite-only)
let dev_config = CompositeConfigBuilder::new()
    .primary("sqlite", BackendKind::Sqlite)
    .build()?;

// Production configuration (PostgreSQL + Elasticsearch)
let prod_config = CompositeConfigBuilder::new()
    .primary("pg", BackendKind::Postgres)
    .search_backend("es", BackendKind::Elasticsearch)
    .sync_mode(SyncMode::Asynchronous)
    .build()?;

Query Routing

Queries are automatically analyzed and routed to optimal backends:

Feature	Detection	Routed To
Basic search	Standard parameters	Primary
Chained parameters	patient.name=Smith	Graph backend
Full-text	_text, _content	Search backend
Terminology	:above, :below, :in	Terminology backend
Writes	All mutations	Primary only
_include/_revinclude	Include directives	Primary

use helios_persistence::composite::{QueryAnalyzer, QueryFeature};
use helios_persistence::types::SearchQuery;

let analyzer = QueryAnalyzer::new();

// Analyze a complex query
let query = SearchQuery::new("Observation")
    .with_parameter(/* _text=cardiac */);

let analysis = analyzer.analyze(&query);
println!("Features: {:?}", analysis.features);
println!("Complexity: {}", analysis.complexity_score);

Result Merging Strategies

When queries span multiple backends, results are merged using configurable strategies:

Strategy	Behavior	Use Case
Intersection	Results must match all backends (AND)	Restrictive queries
Union	Results from any backend (OR)	Inclusive queries
PrimaryEnriched	Primary results with metadata from secondaries	Standard search
SecondaryFiltered	Filter secondary results through primary	Search-heavy queries

Synchronization Modes

Mode	Latency	Consistency	Use Case
Synchronous	Higher	Strong	Critical data requiring consistency
Asynchronous	Lower	Eventual	Read-heavy workloads
Hybrid	Balanced	Configurable	Search indexes sync, others async

use helios_persistence::composite::SyncMode;

// Synchronous: All secondaries updated in same transaction
let sync = SyncMode::Synchronous;

// Asynchronous: Update via event stream
let async_mode = SyncMode::Asynchronous;

// Hybrid: Sync for search indexes, async for others
let hybrid = SyncMode::Hybrid { sync_for_search: true };

Cost-Based Optimization

The cost estimator uses benchmark-derived costs to make routing decisions:

use helios_persistence::composite::{CostEstimator, QueryCost};
use helios_persistence::types::SearchQuery;

let estimator = CostEstimator::with_defaults();
let query = SearchQuery::new("Patient");

// Estimate cost for each backend
let costs = estimator.estimate_all(&query, &config);
for (backend_id, cost) in costs {
    println!("{}: total={}, latency={}ms",
        backend_id, cost.total, cost.estimated_latency_ms);
}

// Get cheapest backend
let best = estimator.cheapest_backend(&query, &config.backends);

Health Monitoring

The health monitor tracks backend availability and triggers failover:

use helios_persistence::composite::{HealthMonitor, HealthConfig};
use std::time::Duration;

let config = HealthConfig {
    check_interval: Duration::from_secs(30),
    timeout: Duration::from_secs(5),
    failure_threshold: 3,  // Mark unhealthy after 3 failures
    success_threshold: 2,  // Mark healthy after 2 successes
};

let monitor = HealthMonitor::new(config);

// Check backend health
if monitor.is_healthy("primary") {
    // Use backend
}

// Get aggregate status
let status = monitor.all_status();
println!("Healthy: {}/{}", status.healthy_count(), status.backends.len());

Configuration Advisor

The configuration advisor is an HTTP API for analyzing and optimizing composite storage configurations.

Running the Advisor

# Build with advisor feature
cargo build -p helios-persistence --features advisor --bin config-advisor

# Run the advisor
./target/debug/config-advisor

# With custom settings
ADVISOR_HOST=0.0.0.0 ADVISOR_PORT=9000 ./target/debug/config-advisor

API Endpoints

Endpoint	Method	Description
/health	GET	Health check
/backends	GET	List available backend types
/backends/{kind}	GET	Get capabilities for a backend type
/analyze	POST	Analyze a configuration
/validate	POST	Validate a configuration
/suggest	POST	Get optimization suggestions
/simulate	POST	Simulate query routing

Example: Analyze Configuration

curl -X POST http://localhost:8081/analyze \
  -H "Content-Type: application/json" \
  -d '{
    "config": {
      "backends": [
        {"id": "primary", "role": "Primary", "kind": "Sqlite"}
      ]
    }
  }'

Example: Get Suggestions

curl -X POST http://localhost:8081/suggest \
  -H "Content-Type: application/json" \
  -d '{
    "config": {
      "backends": [
        {"id": "primary", "role": "Primary", "kind": "Sqlite"}
      ]
    },
    "workload": {
      "read_ratio": 0.8,
      "write_ratio": 0.2,
      "fulltext_search_ratio": 0.3,
      "queries_per_day": 10000
    }
  }'

Example Configurations

Development (SQLite-only)

let config = CompositeConfigBuilder::new()
    .primary("sqlite", BackendKind::Sqlite)
    .build()?;

Production with Full-Text Search

let config = CompositeConfigBuilder::new()
    .primary("pg", BackendKind::Postgres)
    .search_backend("es", BackendKind::Elasticsearch)
    .sync_mode(SyncMode::Asynchronous)
    .build()?;

Graph-Heavy Workloads

let config = CompositeConfigBuilder::new()
    .primary("pg", BackendKind::Postgres)
    .graph_backend("neo4j", BackendKind::Neo4j)
    .sync_mode(SyncMode::Hybrid { sync_for_search: false })
    .build()?;

Large-Scale Archival

let config = CompositeConfigBuilder::new()
    .primary("s3", BackendKind::S3)
    .search_backend("es", BackendKind::Elasticsearch)
    .sync_mode(SyncMode::Synchronous)
    .build()?;

Troubleshooting

Query not routing to expected backend:

• Enable debug logging: RUST_LOG=helios_persistence::composite=debug
• Use the analyzer to inspect detected features: analyzer.analyze(&query)
• Check backend capabilities match required features

High sync lag:

• Reduce batch size in SyncConfig
• Increase sync workers
• Consider synchronous mode for critical data

Failover not triggering:

• Check health check interval isn't too long
• Verify failure threshold is appropriate
• Ensure failover_to targets are configured

Cost estimates seem wrong:

• Run Criterion benchmarks to calibrate costs
• Use with_benchmarks() on CostEstimator
• Check feature multipliers in CostConfig

License

MIT