You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In src/otel/metrics.rs, the insert_exemplars helper (introduced in PR #1661 to replace the old flatten_exemplar) iterates over the Vec<Exemplar> for a data point and writes each exemplar's fields directly into the same parent map, using static key names:
exemplar_time_unix_nano
exemplar_span_id
exemplar_trace_id
exemplar_value
Because every exemplar in the slice writes to the same keys, each iteration overwrites the previous one. After the loop only the last exemplar is retained in the flattened record. All prior exemplars are silently dropped.
This affects Sum, Histogram, and ExponentialHistogram metric types — all of which carry exemplars as Vec<Exemplar>.
The previous implementation (flatten_exemplar) returned a Vec<Map<String, Value>> (one Map per exemplar), which at least preserved each exemplar in a separate structure, though the call sites then merged only the first map.
Ideal Behavior
Exemplars for a data point should be preserved in full. Possible approaches:
Array column – Serialize the exemplars as a Value::Array of Value::Object entries and store them under a single column (e.g., exemplars). Each element would contain exemplar_time_unix_nano, exemplar_span_id, exemplar_trace_id, exemplar_value, and any filtered attributes. This is the most semantically faithful representation and aligns with how the OTel spec treats exemplars (a list per data point).
Indexed flat columns – Emit exemplar_0_time_unix_nano, exemplar_1_time_unix_nano, … etc. This avoids nested types but creates a variable, unbounded number of columns.
Emit one row per exemplar – Fan out exemplar data so each exemplar becomes its own row joined to its parent data point's fields. This is queryable without nested-type support but multiplies row count.
Option 1 (array column) is recommended as it matches the OTel data model, keeps schema stable, and allows query-layer unnesting when needed.
Summary
Tracked from PR #1661 (comment: #1661 (comment)).
Reported by @parmesant.
Current Behavior
In
src/otel/metrics.rs, theinsert_exemplarshelper (introduced in PR #1661 to replace the oldflatten_exemplar) iterates over theVec<Exemplar>for a data point and writes each exemplar's fields directly into the same parent map, using static key names:exemplar_time_unix_nanoexemplar_span_idexemplar_trace_idexemplar_valueBecause every exemplar in the slice writes to the same keys, each iteration overwrites the previous one. After the loop only the last exemplar is retained in the flattened record. All prior exemplars are silently dropped.
This affects Sum, Histogram, and ExponentialHistogram metric types — all of which carry exemplars as
Vec<Exemplar>.The previous implementation (
flatten_exemplar) returned aVec<Map<String, Value>>(oneMapper exemplar), which at least preserved each exemplar in a separate structure, though the call sites then merged only the first map.Ideal Behavior
Exemplars for a data point should be preserved in full. Possible approaches:
Array column – Serialize the exemplars as a
Value::ArrayofValue::Objectentries and store them under a single column (e.g.,exemplars). Each element would containexemplar_time_unix_nano,exemplar_span_id,exemplar_trace_id,exemplar_value, and any filtered attributes. This is the most semantically faithful representation and aligns with how the OTel spec treats exemplars (a list per data point).Indexed flat columns – Emit
exemplar_0_time_unix_nano,exemplar_1_time_unix_nano, … etc. This avoids nested types but creates a variable, unbounded number of columns.Emit one row per exemplar – Fan out exemplar data so each exemplar becomes its own row joined to its parent data point's fields. This is queryable without nested-type support but multiplies row count.
Option 1 (array column) is recommended as it matches the OTel data model, keeps schema stable, and allows query-layer unnesting when needed.
References
repeated ExemplaronNumberDataPoint,HistogramDataPoint, andExponentialHistogramDataPoint.