Missing type casting in AGG group qualifier encoding

[zipdoki]

Bug

AGG queries return count=0 when a group field's property type is non-String (e.g., Long, Int).

Cause

Write path (mutation): Group field values are serialized using their native runtime type. For example, 1L (Long) is encoded via encodeInt64() (header 0x2C).

Read path (AGG query): WherePredicate.parse(ranges) parses the query string without schema context, so "someLongField:eq:1" yields a String "1". This value reaches ValueUtils.serialize() without any type casting, and is encoded via encodeString() (header 0x34).

The qualifier bytes differ, so the HBase lookup returns no match.

The scan path avoids this problem by applying StartStopItem.ensureType() to cast predicate values to the correct schema type before encoding. The AGG path needs an equivalent type casting step.

Note: String-typed fields happen to work because both paths produce the same encoding. Fields with a bucket also work because bucket.handleQueryValue().toString() normalizes the value on both paths. These cases mask the problem.

Steps to reproduce

1. Define a group with a non-String field:

mapOf(
    "group" to "my_group",
    "type" to "COUNT",
    "fields" to listOf(
        mapOf("name" to "someLongField", "bucket" to null),
        mapOf("name" to "timestamp", "bucket" to mapOf(/* date bucket */)),
    ),
)

2. Insert edges with someLongField = 1L
3. Query: GET .../edges/agg/my_group?...&ranges=someLongField:eq:1;time:eq:...

Expected: Returns the aggregated count
Actual: Returns count=0

Suggested fixes

Option A: Cast predicate values before encoding (minimal change)

Follow the same pattern as scan's StartStopItem.ensureType()

val eqValues = firstPairs.map { (field, predicate) ->
    require(predicate is WherePredicate.Eq)
    val castedValue = if (field.bucket == null) {
        schema.getField(field.name).type.cast(predicate.value)
    } else {
        predicate.value
    }
    field.bucketOrGet(castedValue, ceil = false)
}

• Pros: Small change, consistent with the existing scan pattern
• Cons: Only fixes V3QueryService.agg(). Future callers of bucketOrGet() remain unprotected.

Option B: Add type information to Group.Field (structural fix)

Make Group.Field type-aware so that bucketOrGet() itself handles casting:

data class Field(
    val name: String,
    val bucket: Bucket? = null,
    val type: FieldType? = null,
) {
    fun bucketOrGet(value: Any, ceil: Boolean): Any =
        bucket?.handleQueryValue(value, ceil)?.toString()
            ?: type?.cast(value)
            ?: value
}

• Pros: Fixes the problem at its source. All callers of bucketOrGet() are safe by default.
• Cons: Wider change — requires propagating type information when constructing Group.Field from the schema.

References

• V3QueryService.agg() — parses ranges without schema
• Group.Field.bucketOrGet() — no type casting when bucket is null
• ValueUtils.serialize() — encodes based on runtime type
• IndexedLabelMixin.StartStopItem.ensureType() — scan's type casting for reference

[em3s]

+1 for Option B. Please open a PR when you're ready. 👍

[zipdoki]

@em3s
Went with Option B. See #235.

[em3s]

@zipdoki
Looking at the PR #235 , I think we need a small tweak to Option B.

type shouldn't live on Field — it should come from the schema. The idea is: if bucket is non-null, return as String; otherwise, cast using the type from schema. So let's move type out of the constructor and into bucketOrGet as a parameter:

data class Field(
    val name: String,
    val bucket: Bucket? = null,
) {
    fun bucketOrGet(value: Any, ceil: Boolean, type: FieldType? = null): Any =
        bucket?.handleQueryValue(value, ceil)?.toString()
            ?: type?.cast(value)
            ?: value
}

[em3s]

@zipdoki
After discussing this over chat, we'd like to go with:

val eqValues = firstPairs.map { (field, predicate) ->
    require(predicate is WherePredicate.Eq)
    if (field.bucket == null) {
        schema.getField(field.name).type.cast(predicate.value)
    } else {
        field.bucketOrGet(predicate.value, ceil = false)
    }
}

Please update or close #235 and open a new PR.