Supported Storage Formats
Supported formats
Tunny Dashboard can read the following sources directly — no conversion is needed.
| Format | Extension / Spec | Description |
|---|---|---|
| Optuna Journal file | .log |
An append-only operation log produced by Optuna's JournalStorage. |
| Optuna SQLite database | .db / .sqlite / .sqlite3 |
A file produced by Optuna's RDBStorage using SQLite. |
| Optuna RDB connection URL | — | PostgreSQL and MySQL backends of Optuna's RDBStorage are read by URL, e.g. postgresql://user:pass@host:5432/db or mysql://user:pass@host:3306/db. |
| DesignExplorer-format CSV | *.csv |
The CSV output format from the DesignExplorer application: one trial per row, with in:<name> parameter columns, out:<name> objective columns, and an optional img column containing relative artifact paths. |
Server database connections support PostgreSQL and MySQL only.
Practical notes
- Files are opened read-only, so you can safely inspect a database that belongs to an optimization still running elsewhere.
- Manual Reload is supported for Journal, SQLite, and PostgreSQL / MySQL sources. Press it to re-read the open storage and pick up new trials on demand.
- DesignExplorer-format CSV is a one-time import with user-confirmed optimization directions and parameter ranges. It does not support Reload.
Loading during optimization
When loading one Study from SQLite, PostgreSQL, or MySQL, related metadata, trial data, user attributes, constraints, and intermediate values are read from one consistent database snapshot, even if another process commits changes during the read. The read transaction lasts only for data acquisition and ends before table conversion or analysis; it is not held while you inspect the results.
The snapshot is established at the first data read, not necessarily when you press Open or Reload. This is not automatic synchronization with subsequent writes: use Reload to acquire later changes. Separate Studies are loaded independently, not as one simultaneous snapshot.
For MySQL, this consistency applies to ordinary Optuna databases using InnoDB tables. It is not a guarantee for other or mixed storage engines. Read-only access does not guarantee that database writers or maintenance will never wait during acquisition.
Recorded values and objective names
Journal loading preserves recorded numeric Infinity, -Infinity, and NaN in
objective values, intermediate values, and user attributes, including nested
attributes. These remain distinct from explicit null and missing values in the
loaded source data. Quoted strings such as "Infinity" remain strings. CSV
columns inferred as numeric user attributes likewise retain nonfinite values
such as inf, -inf, and NaN.
Preserving these observations does not mean every chart or numerical method
supports infinite inputs: existing method-specific finite-input filtering and
analysis policies are unchanged. JSON-only output cannot represent nonfinite
numbers faithfully and may normalize them to null; it should not be used to
infer that the source recorded null or no value. This Journal behavior does
not imply that every database backend accepts nonfinite user attributes.
If metric names occur later in a Journal, they replace the earlier objective labels without adding another set of objective columns or changing the parameter column order. Objectives with repeated labels keep separate values according to their objective index; the label alone does not identify an objective.
Multi-objective report candidates
Reports select COMPLETE trials with all objective values finite before
computing nondominance, using each objective's minimize/maximize direction.
Trials with Infinity, -Infinity, or NaN in any objective do not
participate in this comparison; their recorded values remain preserved in the
loaded data.
- Unconstrained Study: the report front is the nondominated set among these finite rows.
- Constrained Study with finite verified Feasible rows: the report front is computed only among those rows. Infeasible and Feasibility unverified trials cannot enter the verified feasible front.
- Constrained Study without finite verified Feasible rows: the report uses objective-only nondominated candidates among the finite rows. This fallback also applies when verified Feasible trials exist but each has at least one nonfinite objective.
- No rows with all objectives finite: there are no front candidates.
Fallback entries are labeled Objective-only candidates (not a verified feasible Pareto front) and retain each trial's actual feasibility label. The report's verified feasible front count is zero in this case. Fallback selection does not change the Study's verified feasible trial count: Feasible trials with nonfinite objectives still count as Feasible, and fallback candidates do not add to that count. Nondominance here is relative to the report's selected finite population, not all recorded trials. These are report candidate-selection rules, not a general finite-input policy for every chart or analysis method.