Killing the Constraint Model
The subscription system had a Constraint(field, op, value) model — structured triples that got AND’d together and compiled to MVEL. It worked for simple cases: status == active, priority IN HIGH,CRITICAL. The issue was what came next.
OR logic? Extend ConstraintOp with compound types. Nesting? Add a recursive group model. Cross-field comparisons like deadline < createdAt + 7d? Now you’re rebuilding an expression language, badly, inside a data model that was never designed for it.
The expression SPI we’d just landed (#141) already had MVEL and JQ engines with lazy compilation, parameterized variable binding, and a concurrent cache. Every capability Constraint would need to grow into already existed one layer down. The structured model was just a translation step that deferred the inevitable.
So we removed it. Subscription now carries List<ExpressionEvaluator> — each entry is a typed expression (MVEL or JQ) that the SubscriptionEngine compiles directly via ExpressionEngineRegistry. The ConstraintCompiler that sat between the two is gone. Constraint and ConstraintOp are deleted from platform-api entirely.
The interesting part was what didn’t change. The alpha network — TypeNode, FilterNode, FilterExpression with collapsed sharing — doesn’t know or care what produced the FilterExpression. Two subscriptions with identical MVEL expressions in the same tenant still share a FilterNode and evaluate the predicate once. The sharing identity is the expression string, not the input model. Swapping Constraint triples for raw expressions didn’t touch a single line in datasource-alpha.
Tenant isolation moved from a hidden concern inside ConstraintCompiler to an explicit wrapping predicate in SubscriptionEngine. The $me owner placeholder became a bound MVEL variable instead of a string substitution — cleaner, and it works natively with MVEL’s compilation model.
One gotcha surfaced during testing: MVEL3 doesn’t support single-quoted string literals. status == 'completed' throws an UnsolvedSymbolException because the transpiler tries to resolve mpleted as a Java identifier. The old compiler never hit this because it used parameterised variables ($p0) instead of string literals. Double quotes work: status == "completed". Non-obvious if you’re coming from MVEL2, which treated both quote styles identically.
The net result is 36 fewer lines of code and one fewer abstraction layer. More importantly, the subscription filter model is now the expression SPI — it doesn’t need to grow its own feature set to handle complex predicates.