[pawellewandowski98]

Load labels and tags by row id to fix "stack depth limit exceeded" on Postgres

What Changed

• Added loadLabels (pkg/internal/storage/repo/transactions.go). FindTransactions uses it instead of GORM's Preload of the many2many Labels.
• Added loadTags (pkg/internal/storage/repo/outputs.go). ListAndCountOutputs (when IncludeTags is set) and FindInputsAndOutputsWithBaskets use it instead of Preload("Tags").
• Both helpers join the link table to bsv_labels / bsv_tags and filter on the single id column: transaction_id IN (...) or output_id IN (...).
• IDs are bound in batches of relationBatchSize (5,000), well under the bind-parameter limits of SQLite (32,766) and Postgres (65,535).
• Preload's soft-delete filters are kept: soft-deleted links, and links to soft-deleted labels or tags, are still left out.
• In ListAndCountOutputs, tags are loaded after the paged query, only for the rows on that page.

Why It Was Necessary

• On Postgres, ListTransactions failed with ERROR: stack depth limit exceeded once a user had a few thousand transactions.
• Label and Tag are keyed by (name, user_id), so Preload loads them with a row-value list, WHERE (name, user_id) IN ((..),(..),...), one tuple per distinct label or tag. Postgres fails to evaluate that list somewhere between 5,000 and 8,000 tuples.
• Wallets that label each transaction, or tag each output, with its own ID reach that quickly. We hit it at ~5.5k transactions on a BCS-Lite stack whose token wallet labels every action with its operation ID.
• The same pattern sat in three read paths:
  • FindTransactions, used by ListTransactions, abort and abort-abandoned;
  • ListAndCountOutputs, used by ListOutputs with tags and by balance;
  • FindInputsAndOutputsWithBaskets, used by ListActions.

Testing Performed

• Added TestFindTransactions_ManyDistinctLabels (repo/transactions_find_labels_test.go):
  • Setup: 9,000 transactions, each with its own label plus a shared one, one soft-deleted link, and one link to a soft-deleted label.
  • Checks that every transaction comes back with exactly its live labels.
• Added TestOutputs_ManyDistinctTags (repo/outputs_tags_test.go):
  • Setup: one transaction with 9,000 outputs, each with its own tag plus a shared one, with the same soft-delete cases.
  • Covers both ListAndCountOutputs with tags and FindInputsAndOutputsWithBaskets.
• Both tests fail on Postgres without the change and pass with it. They also pass on SQLite:
  • go test ./pkg/internal/storage/repo/ -run ManyDistinct
  • TEST_DB_MODE=postgres go test ./pkg/internal/storage/repo/ -run ManyDistinct

Impact / Risk

• No API, schema or migration changes. The returned entities don't change shape.
• Regression risk is low. Only how labels and tags are loaded changes, and the new tests cover the soft-delete cases.
• Performance:
  • Labels and tags now take one extra query per 5,000 rows, instead of the Preload query.
  • For the page-sized results that ListOutputs and ListActions return, the cost is about the same.
  • Large result sets no longer fail.
• Not changed:
  • The remaining Preload("Labels") calls load a single transaction, so they never build a large list.
  • The row-value (tx_id, vout) IN (...) list for no-send change outpoints stays small in practice.

[sonarqubecloud]

Quality Gate passed

Issues
1 New issue
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud