explain-plan

You are helping the user understand and optimize query execution plans in Apache DataFusion.

Input: $@

Step 1 — Check datafusion-cli is installed

command -v datafusion-cli

If not found, delegate to /datafusion-skills:install-datafusion.

Step 2 — Resolve state

STATE_DIR=""
test -f .datafusion-skills/state.sql && STATE_DIR=".datafusion-skills"
PROJECT_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || echo "$PWD")"
PROJECT_ID="$(echo "$PROJECT_ROOT" | tr '/' '-')"
test -f "$HOME/.datafusion-skills/$PROJECT_ID/state.sql" && STATE_DIR="$HOME/.datafusion-skills/$PROJECT_ID"

Step 3 — Determine the mode

• If --analyze is present → use EXPLAIN ANALYZE (actually runs the query, shows real metrics)
• Otherwise → use EXPLAIN (shows the plan without execution)

Extract the SQL query (remove --analyze flag if present).

If the input is natural language, generate SQL first (see /datafusion-skills:query for SQL generation guidelines).

Step 4 — Run EXPLAIN

Physical plan (default — shows the execution plan as a visual tree):

datafusion-cli ${STATE_DIR:+--file "$STATE_DIR/state.sql"} -c "
EXPLAIN $SQL;
"

Verbose plan (full optimizer trace — initial logical plan, each optimization pass, initial physical plan, final physical plan with stats and schema):

datafusion-cli ${STATE_DIR:+--file "$STATE_DIR/state.sql"} -c "
EXPLAIN VERBOSE $SQL;
"

With actual metrics (if --analyze) (runs the query, reports per-operator row counts, timing, memory, spill stats):

datafusion-cli ${STATE_DIR:+--file "$STATE_DIR/state.sql"} -c "
EXPLAIN ANALYZE $SQL;
"

Step 5 — Analyze the plan

Parse the execution plan output and provide insights:

Key things to look for:

1. Full table scans → Look for TableScan without pushdown predicates
   • Suggest adding WHERE clauses or partitioning
   • Check if filter pushdown is happening
2. Sort operations → SortExec or SortPreservingMergeExec
   • Expensive for large datasets
   • Suggest pre-sorting data or using sorted Parquet files
3. Hash joins vs merge joins → HashJoinExec vs SortMergeJoinExec
   • Hash joins need memory for the build side
   • Suggest which table should be the build side (smaller table)
4. Repartitioning → RepartitionExec
   • Shows data shuffling between partitions
   • Can be expensive for large datasets
5. Projection pushdown → Check if only needed columns are read
   • DataFusion should push projections down to the scan
6. Predicate pushdown → Check if filters are pushed to the scan level
   • Look for predicate in TableScan nodes
7. Coalesce partitions → CoalescePartitionsExec
   • Merging partitions back to single partition
   • Expected at the top of the plan

For EXPLAIN ANALYZE, additionally check:

• Row counts at each stage → identify data amplification or reduction
• Execution time per operator → find the bottleneck
• Memory usage → identify memory-intensive operations

Step 6 — Present findings

Structure the analysis as:

Query Plan Summary

Brief description of what the plan does.

Plan Visualization

Present the plan as an indented tree (already DataFusion's default output format).

Performance Analysis

• Bottlenecks: Operations that are likely slowest
• Optimizations applied: Filter pushdown, projection pushdown, etc.
• Opportunities: Suggestions for improving performance

Recommendations

Actionable suggestions, such as:

• Add indexes or sort data
• Rewrite the query to enable better pushdown
• Adjust DataFusion configuration options
• Use partitioned data layout

Step 7 — Suggest configuration tuning

If relevant, suggest DataFusion configuration changes:

-- Increase target partitions for more parallelism
SET datafusion.execution.target_partitions = 8;

-- Increase batch size for throughput
SET datafusion.execution.batch_size = 16384;

-- Enable/disable optimizations
SET datafusion.optimizer.enable_round_robin_repartition = true;

To explore these settings, try /datafusion-skills:datafusion-docs configuration options.