Opik Backend

Architecture

• Layered: Resource → Service → DAO (never skip layers)
• DI: Guice modules, constructor injection with @Inject
• Databases: MySQL (metadata, transactional) + ClickHouse (analytics, append-only)

Naming Conventions

Plural Names (Resources, Tests, URLs, DB Tables)

• Resource classes: TracesResource, SpansResource, DatasetsResource (not TraceResource)
• Resource test classes: TracesResourceTest, SpansResourceTest, DatasetsResourceTest (not TraceResourceTest)
• URL paths: /v1/private/traces, /v1/private/spans (not /v1/private/trace)
• DB table names: traces, spans, feedback_scores (not trace, span, feedback_score)

Singular Names (DAO, Service)

• DAO classes: TraceDAO, SpanDAO, DatasetDAO (not TracesDAO)
• Service classes: TraceService, SpanService, DatasetService (not TracesService)

// ✅ GOOD
@Path("/v1/private/traces")
public class TracesResource { }

// ✅ GOOD - DAO and Service use singular
public class TraceDAO { }
public class TraceService { }

// ✅ GOOD - test classes match plural resource name
public class TracesResourceTest { }

// ❌ BAD - singular test class
public class TraceResourceTest { }

// ❌ BAD - singular resource/URL
@Path("/v1/private/trace")
public class TraceResource { }

// ❌ BAD - plural DAO/Service
public class TracesDAO { }
public class TracesService { }

Lombok Conventions

Records and DTOs

• Always annotate records/DTOs with @Builder(toBuilder = true)
• Use builders (not constructors) when instantiating records
• For internal records (built programmatically, never validated by Bean Validation), use Lombok @NonNull on required fields — it generates a runtime null check at construction
• For request-body DTOs validated via @Valid cascade (Jakarta validators like @NotNull/@NotBlank/@Size), use Jakarta annotations only — do not stack @NonNull on top. Bean Validation already enforces the contract at the API boundary; doubling up is redundant noise

// ✅ GOOD - internal record, Lombok @NonNull
@Builder(toBuilder = true)
record MyData(@NonNull UUID id, @NonNull String name, String description) {}

MyData data = MyData.builder()
        .id(id)
        .name(name)
        .build();

// ✅ GOOD - request-body DTO, Jakarta validators only
@Builder(toBuilder = true)
public record MyRequest(
        @NotNull UUID id,
        @NotBlank String name,
        @NotNull @Size(min = 1, max = 1000) @Valid List<MyItem> items) {}

// ❌ BAD - plain constructor (positional mistakes, less readable)
new MyData(id, name, null);

// ❌ BAD - @Builder without toBuilder
@Builder
record MyData(UUID id, String name) {}

// ❌ BAD - stacking @NonNull and @NotNull on the same field
public record MyRequest(@NonNull @NotNull UUID id) {}

Dependency Injection

• Use @RequiredArgsConstructor(onConstructor_ = @Inject) instead of manual constructors

// ✅ GOOD
@RequiredArgsConstructor(onConstructor_ = @Inject)
public class MyService {
    private final @NonNull DependencyA depA;
    private final @NonNull DependencyB depB;
}

// ❌ BAD - boilerplate constructor
public class MyService {
    private final DependencyA depA;
    @Inject
    public MyService(DependencyA depA) {
        this.depA = depA;
    }
}

Interfaces

• Don't put validation annotations (@NonNull) on interface method parameters
• Keep interfaces free of implementation details

// ✅ GOOD
interface MyService {
    void process(String workspaceId, UUID promptId);
}

// ❌ BAD - validation on interface
interface MyService {
    void process(@NonNull String workspaceId, @NonNull UUID promptId);
}

Critical Gotchas

StringTemplate Memory Leak

// ✅ GOOD
var template = TemplateUtils.newST(QUERY);

// ❌ BAD - causes memory leak via STGroup singleton
var template = new ST(QUERY);

List Access

// ✅ GOOD
users.getFirst()
users.getLast()

// ❌ BAD
users.get(0)
users.get(users.size() - 1)

SQL Query Construction

Never build a query out of Java string operations. No +, no String.format / .formatted(...), no StringBuilder, no MessageFormat, no String.join over clauses. A query is declared once as a text block, and everything that varies goes through exactly one of two mechanisms:

What varies	Mechanism
A value — id, name, timestamp, list of ids	:placeholder + .bind("placeholder", value)
A fragment — predicate, sort clause, projected column, CTE	StringTemplate <if(x)>…<endif>, <else>, <x> + template.add("x", …)

Why: interpolating values is the SQL-injection surface, and interpolating fragments hides which query a DAO actually runs — the declaration site stops being readable, and callers drift apart over time.

// ✅ GOOD - text block, values bound, structure via StringTemplate
@SqlQuery("""
        SELECT * FROM datasets
        WHERE workspace_id = :workspace_id
        <if(name)> AND name like concat('%', :name, '%') <endif>
        """)

// ❌ BAD - string concatenation
@SqlQuery("SELECT * FROM datasets " +
        "WHERE workspace_id = :workspace_id " +
        "<if(name)> AND name like concat('%', :name, '%') <endif> ")

A predicate that differs between callers is a fragment, so it belongs in the template — not in a %s slot the caller fills in:

// ❌ BAD - caller splices the predicate in
private static final String TOKEN_USAGE_NAMES_TEMPLATE = """
        SELECT DISTINCT name FROM (
            SELECT usage FROM spans FINAL
            WHERE workspace_id = :workspace_id
            AND %s
        ) ...
        """;

static String tokenUsageNames(String projectPredicate) {
    return TOKEN_USAGE_NAMES_TEMPLATE.formatted(projectPredicate);
}

// caller: tokenUsageNames("project_id IN :project_ids")

// ✅ GOOD - both shapes live in the template, the caller picks one
private static final String TOKEN_USAGE_NAMES = """
        SELECT DISTINCT name FROM (
            SELECT usage FROM spans FINAL
            WHERE workspace_id = :workspace_id
            <if(project_ids)> AND project_id IN :project_ids <endif>
            <if(project_id)> AND project_id = :project_id <endif>
        ) ...
        """;

var template = TemplateUtils.newST(TOKEN_USAGE_NAMES);
template.add("project_ids", true);
...
statement.bind("project_ids", projectIds.toArray(new UUID[0]));

A fragment that genuinely can't be enumerated in the template — a user-chosen sort field or filter clause — must be produced by the allow-listed builders (SortingQueryBuilder, FilterQueryBuilder), never assembled from raw request strings.

.formatted(...) stays correct for log and exception messages. The rule is about SQL text only.

Some %s query templates predate this rule. Don't copy them and don't add new ones.

Immutable Collections

// ✅ GOOD
Set.of("A", "B", "C")
List.of(1, 2, 3)
Map.of("key", "value")

// ❌ BAD
Arrays.asList("A", "B", "C")

API Design

• Query parameters that accept lists: Use plural names from the start (e.g., exclude_category_names not exclude_category_name). Starting with a singular name and later adding a plural variant results in two redundant query params on the same endpoint. Plural names are backward-compatible since they work for both single and multiple values.

Error Handling

Use Jakarta Exceptions

throw new BadRequestException("Invalid input");
throw new NotFoundException("User not found: '%s'".formatted(id));
throw new ConflictException("Already exists");
throw new InternalServerErrorException("System error", cause);

Error Response Classes

• Simple: io.dropwizard.jersey.errors.ErrorMessage
• Complex: com.comet.opik.api.error.ErrorMessage
• Never create new error message classes

Logging

Format Convention

// ✅ GOOD - values in single quotes
log.info("Created user: '{}'", userId);
log.error("Failed for workspace: '{}'", workspaceId, exception);

// ❌ BAD - no quotes
log.info("Created user: {}", userId);

Never Log

• Emails, passwords, tokens, API keys
• PII, personal identifiers
• Database credentials

Reference Files

• clickhouse.md - ClickHouse query patterns
• mysql.md - TransactionTemplate patterns
• testing.md - PODAM, naming, assertion patterns
• migrations.md - Liquibase format for MySQL/ClickHouse
• permissions.md - @RequiredPermissions annotation guidance for endpoints