dataxplan 0.1.1

Read PostgreSQL EXPLAIN plans locally: parse the plan, compute self time and estimation error, flag documented problems, compare plans, and guard them in CI. No database connection, deterministic.

dataxplan

Read PostgreSQL EXPLAIN plans from Python: parse the plan, compute the numbers people misread (self time and estimation error), flag documented problems, compare plans, and guard them in CI. No database connection, nothing leaves your machine, deterministic output.

You give it the output of EXPLAIN (ANALYZE, BUFFERS, FORMAT JSON) ...; it does the rest locally.

Problem

A slow query's only honest explanation is in its EXPLAIN plan, but reading that plan by hand is hard and easy to get wrong. The reported times are per loop and inclusive of children, so the node that is actually slow is rarely the obvious one, and the row estimate that is off by orders of magnitude (the usual root cause of a bad plan) is buried in the output. The tools that read plans well are web pastebins, which means sending a production plan, with its table names and filter values, to someone else's server, or commercial SaaS. None of them run inside your own script, your tests, or your CI.

Purpose

dataxplan reads the plan for you, locally and deterministically. From the output of EXPLAIN (ANALYZE, BUFFERS, FORMAT JSON) it builds the plan tree, computes the numbers people misread (self time, estimation error, disk spills), flags documented problems with an explanation and a suggested action, compares two plans for regression, and lets you assert on a plan in CI so a change that degrades it fails the build. It needs no database connection, has no runtime dependencies, and every result is reproducible. The findings are documented heuristics, not guarantees, and the analysis is of the plan you provide.

Methodology

The metrics are computed directly from the fields PostgreSQL documents for EXPLAIN (ANALYZE, BUFFERS, FORMAT JSON), so they are arithmetic on the plan, not numbers of our own. For a node n with children C(n):

• de-looped total (inclusive) time: total(n) = Actual Total Time(n) × Actual Loops(n)
• self (exclusive) time, the work done at the node itself: self(n) = total(n) − Σ total(c) for c in C(n)
• share of the query's time: pct(n) = self(n) / Execution Time
• row estimation error, per loop (both counts floored at 1): factor(n) = Actual Rows(n) / Plan Rows(n) and error(n) = max(factor(n), 1 / factor(n))
• a sort or hash spilled to disk when Sort Method contains external, or Hash Batches > 1, or Temp Written Blocks > 0.

Actual Total Time is reported per loop and inclusive of children, which is why both the de-looping (× Actual Loops) and subtracting the children matter: together they reveal the node where the time is actually spent, not the one at the top of the plan. The findings are documented heuristics applied on top of these metrics (sources in References); they are pointers, not guarantees.

It turns a plan into a deterministic read. Here the Bosch manufacturing example (examples/): a hash join over a 48-million-row scan of the production-line measurements, estimated at 300 rows but producing 288,000 (Bosch Production Line Performance):

dataxplan
  execution time   2,601.00 ms   (planning 0.50 ms)
  nodes 5, depth 3
  worst row estimate   960x off
  spilled to disk      no
  top by self time:
    Seq Scan on measurements         2,200.00 ms (85%)
    Hash Join                        320.00 ms (12%)
    Seq Scan on parts                55.00 ms (2%)
    Aggregate                        20.00 ms (1%)
    Hash                             5.00 ms (0%)
findings:
  [HIGH] Hot sequential scan  (Seq Scan on measurements)
      sequential scan is 85% of execution time, reading 48,000,000 rows
      -> consider an index supporting the filter or join on measurements
  [HIGH] Row estimate is far off  (Hash Join)
      estimated 300 rows, actual 288,000 (960x under-estimate)
      -> run ANALYZE on the table; if the columns are correlated consider extended statistics
  [MEDIUM] Filter discards most rows read  (Seq Scan on parts)
      removed 1,194,000 rows by filter but kept only 6,000
      -> the predicate is not selective via the current access path; an index may help

Install

pip install dataxplan

No runtime dependencies. The chart is optional (pip install "dataxplan[viz]").

Quick start

import dataxplan

report = dataxplan.analyze(explain_json)   # the EXPLAIN (FORMAT JSON) output
print(report.summary())                    # the summary shown above

From the command line

dataxplan plan.json                       # summary
dataxplan plan.json --tree                # also the annotated plan tree
dataxplan plan.json --json                # the full report as JSON
dataxplan before.json --compare after.json
psql -XqAt -c "EXPLAIN (ANALYZE, BUFFERS, FORMAT JSON) <query>" | dataxplan

Guard a plan in CI

Pin a critical query's plan in your test suite, so a code or schema change that makes it regress fails the build. Nothing else in Python does this.

def test_orders_lookup_stays_fast():
    report = dataxplan.analyze(get_explain("SELECT * FROM orders WHERE customer_id = %s"))
    assert not report.has_seq_scan_on("orders")
    assert report.max_estimation_error < 100
    assert not report.spilled_to_disk

Compare two plans (before / after an index)

print(dataxplan.compare(before_json, after_json).summary())
# dataxplan compare - IMPROVED
#   execution time   905.00 ms -> 0.08 ms (-100%)
#   resolved         filter_discard, seq_scan_hot

Sharper findings with catalog context (optional)

from dataxplan import Context, TableInfo
ctx = Context(tables={"orders": TableInfo("orders", row_count=10_000_000,
                                          indexed_columns=("id",))})
dataxplan.analyze(explain_json, context=ctx)

Fetch a plan from a connection you already have (optional)

plan = dataxplan.run_explain(conn, "SELECT * FROM orders WHERE id = %s", params=(42,))
dataxplan.analyze(plan)

run_explain calls cursor.execute on a DB-API connection you pass (psycopg, psycopg2, ...); dataxplan does not depend on any driver. With analyze=True it runs the query, so use analyze=False for a plan-only estimate.

What it covers

Area	What you get
Parse	parse -> a typed Plan / PlanNode tree from EXPLAIN (FORMAT JSON)
Metrics	self (exclusive) time, % of total, estimation error, disk spills, buffers
Findings	hot sequential scans, large row mis-estimates, disk spills, filter discards, nested-loop blow-ups, index-only heap fetches, lossy bitmaps, JIT overhead
Report	summary, to_dict, and an assertion API (has_seq_scan_on, max_estimation_error, spilled_to_disk, ok) for CI
Compare	compare two plans for regression (timing, shape, estimates, findings)
Render	text_tree (annotated, dependency-free) and plan_tree_chart (needs matplotlib)
Context	optional catalog metadata (sizes, indexes, stale stats) that sharpens findings
CLI	dataxplan plan.json (or stdin): summary, --tree, --json, --compare

Examples

Seven plans from public datasets and benchmarks (in examples/), ordered here from the largest data set to the smallest. Each chart shows self time per node, with the high-severity findings in a warning colour. The benchmark and NYC sets are large in their own right; the semiconductor, automotive and textile sets are smaller research samples of domains that run at production scale, so those plans model a production-size query. Sizes and times are illustrative.

NYC TLC Yellow Taxi - a sort that spills to disk

The NYC TLC trip records are a public, multi-billion-row data set. Ordering long trips by fare, dataxplan flags disk_spill (an external-merge sort) over a hot sequential scan (seq_scan_hot).

The red bar is the sequential scan feeding the sort, and the sort then overflowed memory to disk. Action: index the filter so the scan shrinks, or raise work_mem (or pre-aggregate) so the sort stays in memory.

TPC-H lineitem - a hot scan discarding most rows

TPC-H is the standard decision-support benchmark, scalable to hundreds of gigabytes. A selective predicate on l_quantity flags seq_scan_hot and filter_discard (59 million rows read, 2 million kept).

Almost all the time is one scan that reads 59 million rows to keep 2 million. Action: add an index on the filtered column so the predicate is served by the index instead of a full scan.

IMDB / Join Order Benchmark - a row mis-estimate

The Join Order Benchmark (Leis et al., VLDB 2015) is the standard cardinality-estimation benchmark. dataxplan flags estimate_off (2 rows estimated, 480,000 actual) and nested_loop_blowup.

The estimate was 2 rows but the join produced 480,000, so the planner chose a nested loop that ran 480,000 times (the inner scan is the red bar). Action: run ANALYZE and add extended statistics so a hash or merge join is chosen.

Bosch Production Line Performance (manufacturing) - a hash join mis-estimate

The Bosch Production Line Performance set is large public manufacturing data. dataxplan flags a hot 48-million-row scan (seq_scan_hot), a 960x join mis-estimate (estimate_off) and a filter discarding most rows (filter_discard).

The 48-million-row measurements scan is the bottleneck (red) and the join estimate is 960x too low. Action: index or partition measurements by part or station, and refresh statistics so the join is planned correctly.

SECOM (semiconductor) - an index-only scan hitting the heap

SECOM (UCI) is sensor data from a semiconductor line. dataxplan flags index_only_heap_fetches (9 million heap fetches, so the table needs a VACUUM) and estimate_off.

The index-only scan should skip the table, but it made 9 million heap fetches because the visibility map is stale. Action: VACUUM the table so the scan skips the heap, and ANALYZE to fix the row estimate.

Mercedes-Benz manufacturing (automotive) - a lossy bitmap scan

The Mercedes-Benz Greener Manufacturing set (Kaggle) is automotive test-bench data. dataxplan flags lossy_bitmap (6 million rows rechecked, so work_mem is too small) and estimate_off.

The bitmap did not fit in work_mem, so it went lossy and rechecked 6 million rows. Action: raise work_mem so the bitmap stays exact, and refresh statistics for the estimate.

Garment factory (textile) - a hot scan discarding most rows

The Productivity Prediction of Garment Employees set (UCI) is garment (textile) manufacturing data. dataxplan flags seq_scan_hot and filter_discard on the production log.

The production_log scan reads 12 million rows to keep 500,000 (red). Action: add an index on the department column so the filter uses the index instead of a full scan.

What is out of scope

dataxplan analyses the plan you give it. By default it does not connect to a database, run your queries, or read your schema, so a finding is a documented heuristic, not a guarantee, and the suggestions are based on the plan alone. It does not rewrite SQL or invent a cost model. It targets PostgreSQL FORMAT JSON output (MySQL may follow).

How the headline metrics work

• Self time. Postgres reports Actual Total Time per loop and inclusive of children, so a node's total is Actual Total Time x Actual Loops, and its self time is that minus the children's totals. Self time is where the work really happens.
• Estimation error. Plan Rows against Actual Rows (per loop); a large ratio is the usual root cause of a bad plan.

References and validation

The metric arithmetic is verified by hand against the semantics PostgreSQL documents (see tests/). Each finding is grounded in the documented behaviour it relies on, written in our own words:

Finding	Source
estimate_off	PostgreSQL: planner statistics, CREATE STATISTICS; Leis et al. (2015)
seq_scan_hot	PostgreSQL: Using EXPLAIN, Indexes
disk_spill	PostgreSQL: work_mem
filter_discard	PostgreSQL: Using EXPLAIN (Rows Removed by Filter)
nested_loop_blowup	PostgreSQL: planner join methods; Leis et al. (2015)
index_only_heap_fetches	PostgreSQL: Index-Only Scans and Covering Indexes
lossy_bitmap	PostgreSQL: work_mem (lossy bitmap recheck)
jit_overhead	PostgreSQL: Just-in-Time Compilation (jit_above_cost)

Primary references:

• PostgreSQL documentation: EXPLAIN and the Using EXPLAIN wiki page.
• V. Leis, A. Gubichev, A. Mirchev, P. Boncz, A. Kemper, T. Neumann, "How Good Are Query Optimizers, Really?", Proceedings of the VLDB Endowment 9(3), 2015 - the study of cardinality mis-estimation behind the Join Order Benchmark, which the estimate_off finding detects (see the JOB example in examples/).

Plan analysis has no single numeric ground truth the way a closed-form formula does, so the claim here is deliberately narrow: the parsing and arithmetic are correct against the documented format, each finding cites the behaviour it relies on, and the suggestions are pointers, not guarantees.

License

MIT. Written and maintained by Atakan Arikan, MSc Student at Tsinghua University and Politecnico di Milano.