Summary

• Fixes a bug where GridstoreError::FlushCancelled (a benign condition during payload index deletion) was converted into a fatal ServiceError instead of being propagated as OperationError::Cancelled
• The Cancelled variant was lost at two intermediate layers that unconditionally wrapped errors with service_error(format!(...)): all 4 mutable index flushers (map, text, numeric, geo) and the struct_payload_index flusher
• After this fix, the cancellation correctly reaches the handler in entry.rs which gracefully skips the flush with a debug log

Root cause

When a payload index is deleted, the gridstore backing it is dropped. A concurrent background flush detects this (weak Arc refs fail to upgrade) and returns GridstoreError::FlushCancelled. There is already:

1. A From<GridstoreError> for OperationError impl that maps FlushCancelled → OperationError::Cancelled
2. A handler in entry.rs that gracefully handles Cancelled by skipping the flush

But the Cancelled variant was destroyed before reaching the handler because:

1. All 4 mutable index flushers used OperationError::service_error(format!(...)) to convert the gridstore error, instead of using the Into<OperationError> conversion
2. struct_payload_index flusher also unconditionally wrapped sub-flusher errors as service_error

This caused the error: "last background flush failed: ... Failed to flush mutable map index gridstore: Flush was cancelled"

Fix

• In the 4 mutable index flushers: convert GridstoreError via Into<OperationError> first, then only re-wrap non-Cancelled variants
• In struct_payload_index flusher: propagate Cancelled errors as-is, matching the pattern already used for vector storage, quantization, and id tracker flushers in entry.rs

Test plan

• cargo check -p segment passes
• CI passes

Made with Cursor