Configure Great Expectations expectations

This skill takes a user from "here is what I want to be true about my data" to a saved suite of expectations that has been run against a real batch, with every check reported individually.

Two objects, and the relationship between them is where the traps live:

• An expectation is a single check with typed parameters — one column, one assertion, one set of bounds.
• An expectation suite is the named, persisted collection of them.

Everything here goes through Great Expectations' public API and produces ordinary project artifacts. Nothing depends on this skill being present afterwards.

The flow

1. Preflight — find out which project (or in-memory session) you are operating on, and tell the user. See references/preflight.md.
2. Check the precondition — a working batch definition must already exist. If none does, hand off and stop.
3. Match what the user described against the shipped expectation catalog. Build only what they described.
4. Register the suite first, then add the matched expectations to it.
5. Validate against a batch and report each expectation's own outcome.
6. Confirm persistence, and offer write-out if the session is in memory.

Steps 2 and 4 are the two places this flow fails silently rather than loudly. Do not reorder them, and do not skip step 5 — a suite that was never run is not a result worth reporting.

What this skill will not do without being asked

• Install, upgrade, or remove a package, or otherwise modify the interpreter, virtual environment, environment variables, or shell state. Not in response to an error, not preemptively while checking whether something is present, and not by announcing an intention to install with a chance to decline attached — none of those is the user asking. When something is missing, name it and hand over the command; running it is a separate act that starts only from the user's own instruction, later.

• Create a project directory on the user's disk. Only after the user has agreed to write the session out and named where.

Step 1 — Preflight

Follow references/preflight.md in full before anything else. It establishes whether you are working against a project on disk or an in-memory session, tells you what to announce to the user, and covers the environment problems that silently masquerade as "no project found".

The outcome you carry forward is a context object and one fact: whether the session is file-backed or in memory. Both are fully supported paths, and the difference matters at step 6.

Step 2 — The precondition: a working batch definition must exist

Expectations are checks against a batch of data. Without a batch definition to retrieve one through, there is nothing to validate against, and there is no acceptable way to improvise data access here. Do not add a data source, do not read a file directly with pandas, do not construct a batch by hand.

Enumerate what the session already has:

def existing_batch_definitions(context):
    """(data source, asset, batch definition) name triples available in this session."""
    found = []
    for datasource in context.data_sources.all().values():
        for asset in datasource.assets:
            for batch_definition in asset.batch_definitions:
                found.append((datasource.name, asset.name, batch_definition.name))
    return found

If this returns an empty list, stop. Tell the user that expectations need a batch definition to run against, that their project (or session) has none yet, and that setting one up is the gx-configure-data-source skill's job. Hand off to it and end this flow — do not carry on and do not offer a workaround. This is a hard stop, not a suggestion.

If it returns more than one, name them and let the user choose; guessing which slice of their data they meant to assert against is not a decision to make for them. Then retrieve the batch:

datasource_name, asset_name, batch_definition_name = existing_batch_definitions(context)[0]
batch_definition = (
    context.data_sources.get(datasource_name)
    .get_asset(asset_name)
    .get_batch_definition(batch_definition_name)
)
batch = batch_definition.get_batch()   # add batch_parameters=... for a partitioned definition

A dataframe asset needs its data supplied at retrieval time — get_batch(batch_parameters={"dataframe": df}) — so ask the user for the dataframe rather than inventing one.

Retrieving a batch is not proof that it works. For SQL assets, retrieval touches nothing. If the batch definition has not been verified in this session, probe it first with batch.head(n_rows=5) inside the wrapper in references/robustness.md, exactly as the data-source flow does. A broken batch definition discovered at validation time reports as a wall of metric errors instead of one clear configuration problem.

If this probe, or validation in step 5, turns up a missing driver or client library, report it and hand over the install command per the standing rule above. Do not install it yourself — not now, before anything has failed, and not later without the user's own go-ahead.

Step 3 — Match what the user described, and only that

Read the shipped catalog rather than working from memory — see references/expectation-catalog.md. It covers locating the index inside the installed package, matching a described check on type, description, and data-quality category, filtering by the backend behind the batch definition, and reading each expectation's parameters from its schema.

Two rules govern this step.

Never invent an expectation type. If nothing in the catalog matches what the user described, say so, present the nearest candidates the catalog actually contains, and offer the custom-expectation path. Report unmatched checks separately as unbuilt, with the reason — never silently drop one, and never substitute a different check that sounds similar. The catalog reference gives the no-match procedure in full.

Build only what the user described. Do not profile the data to find things worth asserting. Do not scan the schema and propose a check per column. Do not append "you might also want" suggestions to the suite. The user decides what is asserted about their data; a check they did not ask for is a guess of yours that will later fail and be read as a real data-quality problem. If they describe something vaguely ("amounts should be reasonable"), ask what bound they mean — do not pick one from the data.

Inspecting a column to inform a question you ask the user is fine. Turning what you observed into an expectation without being asked is not.

Step 4 — Register the suite first, then add expectations

Register the suite with the context before adding any expectation to it. This is the ordering rule, and getting it backwards loses work silently:

import great_expectations as gx

SUITE_NAME = "orders_quality"

# Reuse an existing suite; create one only when the name is genuinely new.
if SUITE_NAME in {suite.name for suite in context.suites.all()}:
    suite = context.suites.get(SUITE_NAME)
else:
    suite = context.suites.add(gx.ExpectationSuite(name=SUITE_NAME))

# Now add. Each add persists immediately, because the suite is registered.
suite.add_expectation(gx.expectations.ExpectColumnValuesToNotBeNull(column="customer"))
suite.add_expectation(gx.expectations.ExpectColumnValuesToBeBetween(column="amount", min_value=0))

Three things this pattern gets right, each verified against real behavior:

• Work on the handle the factory returns. context.suites.add(...) re-reads the stored suite and returns a different object from the one you passed in. The returned handle is the one wired to the store, so every add_expectation on it is written through immediately, with no separate save call.
• An unregistered suite persists nothing. Building gx.ExpectationSuite(name=...), adding expectations to it, and validating with it all work — batch.validate(suite) runs happily against a suite the context has never seen. Nothing raises, nothing warns, and afterwards context.suites.get(name) raises DataContextError: ExpectationSuite with name <name> was not found. The entire suite is gone. This is the failure mode the ordering rule exists to prevent.
• context.suites.add_or_update() is destructive against an existing name. It replaces the stored suite wholesale rather than merging: calling context.suites.add_or_update(gx.ExpectationSuite(name=SUITE_NAME)) against a suite already holding three expectations leaves it holding zero — no error, no warning. Use the fetch-first pattern above instead. Reach for add_or_update only when the user has asked to replace a suite's whole contents, and tell them what is being discarded first.

Reusing a suite is safe: fetching it returns its existing expectations intact, and adding to it appends. Adding an expectation identical to one already present is a no-op — the collection is set-like — so re-running the flow does not accumulate duplicates.

Suite names: no dots

Never put a dot in a suite name. The store treats dots as path separators. A suite named orders.quality is written to gx/expectations/orders/quality.json, and a.b.c to gx/expectations/a/b/c.json — one nested directory per segment. The suite still loads by its full dotted name, so nothing appears broken from the API, but the project's expectations directory fans out into a tree where nobody looking for a suite file will find it, and a plain listing of the directory no longer shows the suites it holds.

Use underscores or hyphens: orders_quality, orders-daily. If the user asks for a dotted name, say why you are changing it rather than changing it silently.

Step 5 — Validate, and report each expectation on its own terms

result = batch.validate(suite)

That is the whole validation call. It builds everything else it needs internally.

Validation is a data-touching operation, so run it inside the duration-tracked wrapper in references/robustness.md — the same one the batch probe uses. A suite over a large table can run for a long time, and the rules there about checking in with the user rather than abandoning a query that keeps running on the platform apply unchanged.

If the run fails because a driver or client library is missing, hand over the install command and stop there — do not install it to get the run passing, and do not announce that you're about to.

Pair results with expectations by configuration, never by position

result.results does not come back in the order the expectations were added. Validation regroups the suite before running it, in two ways. Both are stable across runs, so neither ever looks like a bug:

• Expectations are grouped by the column they address. Every check on one column is evaluated together, in the order that column was first mentioned. A suite added as not_be_null(customer), mean_to_be_between(amount), values_to_be_unique(customer), max_to_be_between(amount) comes back with the two customer checks first and the two amount checks after. Checks with no column argument at all, such as a table row count, form a group of their own.
• An expectation whose metric errored is moved ahead of every expectation that ran, so a broken parameter also changes the position of everything else.

What makes this dangerous is that plenty of suites do come back in the order they were built: any suite written column by column already matches the grouping, and so does a two- or three-expectation suite over one column. Pairing your input list with the results by index therefore looks correct while you are trying it out and mislabels every finding once the suite grows.

Read the identity off each result instead: each.expectation_config.type and each.expectation_config.kwargs.

Separate a metric error from a data failure

success is False means two very different things, and reporting them the same way tells the user their data is bad when their configuration is:

for each in result.results:
    config = each.expectation_config
    if each.success:
        print(f"PASS {config.type} {config.kwargs}")
    elif not each.result:
        # The metric never evaluated — a configuration problem, not a finding.
        for _metric_id, info in each.exception_info.items():
            print(f"ERROR {config.type} {config.kwargs}: {info['exception_message']}")
    else:
        # The data genuinely failed the check.
        print(f"FAIL {config.type} {config.kwargs}: {each.result}")

Both halves of not each.success and not each.result are load-bearing. A passing expect_column_to_exist also carries an empty result dict, so the emptiness test alone would misclassify it.

On the error branch, exception_info already names the cause exactly — do not guess at it. It is a dict with one entry per metric that errored, keyed by the string repr of a metric identifier, and each value carries an exception_message holding the real cause verbatim: a mistyped column reports Error: The column "nope" in BatchData does not exist. Relay that message as the why, apply the length ceiling from references/robustness.md, and never paste the accompanying exception_traceback.

Never report an empty-result failure as "your data failed this check." It was never evaluated. Report it as an expectation that could not run, name the cause, and offer to fix the parameter and re-run.

Degenerate data does not raise

Empty tables and all-null columns are ordinary results, not exceptions. Anything in your report that treats them as errors is wrong. Verified behavior, on a SQL batch:

Two of these mislead if reported literally:

• {"observed_value": None} is a populated result, so the discriminator above correctly classifies it as a data failure rather than a metric error. But it does not mean the mean was out of range — it means there were no non-null values to compute over. Say that, and say which column.
• A value-level check passing over an all-null column is not reassurance. Nulls are excluded from the unexpected count, so a range check over a column of nothing but nulls succeeds. If the user's real question was whether the column has usable data, pair it with a non-null check and say why.

Summarize honestly

result.describe() returns a JSON string carrying overall success, a statistics block (evaluated_expectations, successful_expectations, unsuccessful_expectations, success_percent), and a per-expectation list of expectation_type / kwargs / success / result. It is a good basis for a written summary — but it does not distinguish a metric error from a data failure. Use the loop above for that distinction and describe() for the counts.

Report, per expectation: what was checked, on which column, and what happened — passed, failed with the observed numbers, or could not run with the cause. Then the totals. Do not present a metric error inside the failure count as though the data had been judged.

Step 6 — Persistence, and write-out when in memory

In a file-backed project the suite is already saved. Registering it in step 4 wrote it to <context_root>/expectations/<suite_name>.json, and every subsequent add_expectation was written through as it happened. There is no save step to run and none to forget. Confirm it concretely: name the suite, name the file, and say that a fresh session picks it up with context.suites.get("<suite_name>").

Validation results are not persisted by this flow. batch.validate() returns the result to you and writes nothing to the project. The suite is the durable artifact; the report you give the user is the record of this run. Say so rather than letting them assume the results are filed somewhere.

In an in-memory session, nothing survives the process. A second in-memory session sees no suites and no data sources at all. Say this plainly — it is a supported way to work, not a degraded one — and offer to write the session out to a real project, per references/write-out.md. That procedure covers the suite along with the data source, asset, and batch definition it depends on; writing out a suite without them leaves a project that cannot run it. Offer it; don't do it unprompted, and don't pick the location.

The offer ends this step, not the flow. Write-out creates a project on the user's disk, so it starts only from their reply — a separate run, after they have agreed and named a directory. Reporting the results and then writing them out in the same breath means they were never asked; putting the write-out call in the same program as the suite and validation calls means the same thing, because there was no point in it where an answer could have arrived. references/write-out.md opens with the gate this depends on. Written out or not, what comes next is below.

Worked example

A file-backed project that already holds a verified batch definition over an orders table. The user asked for two things: customer must never be missing, and amounts must never be negative.

import great_expectations as gx

context = gx.get_context(cloud_mode=False)          # step 1, per references/preflight.md

batch_definition = (                                 # step 2, from what already exists
    context.data_sources.get("warehouse")
    .get_asset("orders")
    .get_batch_definition("all_rows")
)
batch = batch_definition.get_batch()

# step 3: both descriptions matched catalog entries; nothing else was invented
if "orders_quality" in {suite.name for suite in context.suites.all()}:
    suite = context.suites.get("orders_quality")     # step 4: register/fetch first
else:
    suite = context.suites.add(gx.ExpectationSuite(name="orders_quality"))

suite.add_expectation(gx.expectations.ExpectColumnValuesToNotBeNull(column="customer"))
suite.add_expectation(gx.expectations.ExpectColumnValuesToBeBetween(column="amount", min_value=0))

result = batch.validate(suite)                       # step 5

for each in result.results:
    config = each.expectation_config
    if each.success:
        print(f"PASS  {config.type} {config.kwargs}")
    elif not each.result:
        for _metric_id, info in each.exception_info.items():
            print(f"ERROR {config.type} {config.kwargs}: {info['exception_message']}")
    else:
        print(f"FAIL  {config.type} {config.kwargs}: {each.result}")

Against a table of four rows where one customer is null and one amount is -5.0, that prints:

FAIL  expect_column_values_to_not_be_null {'batch_id': 'warehouse-orders', 'column': 'customer'}: {'element_count': 4, 'unexpected_count': 1, 'unexpected_percent': 25.0, 'partial_unexpected_list': [None], ...}
FAIL  expect_column_values_to_be_between {'batch_id': 'warehouse-orders', 'column': 'amount', 'min_value': 0.0}: {'element_count': 4, 'unexpected_count': 1, 'unexpected_percent': 25.0, 'partial_unexpected_list': [-5.0], ...}

This two-expectation suite happens to come back in the order it was built — which is exactly why the loop reads each result's own expectation_config rather than trusting the position. Report it to the user as one of four rows has no customer, and one amount is negative (-5.0); the partial_unexpected_list is what makes a finding actionable, so relay it.

Where this flow ends

The saved, run, reported suite is this flow's own end state. The one thing that stays outside it is rendering results anywhere but this conversation — report what the run found; do not build reporting surfaces the user did not ask for. Setting up data access when the precondition fails (step 2) and building only what the user described rather than profiling (step 3) are already hard rules earlier in this flow, not items this closing note needs to repeat.

What no longer ends here: turning a validated suite into a check that survives the session. Offer the gx-configure-checkpoint skill as the next step — it binds the batch definition and suite this flow just built into a persisted, re-runnable checkpoint with post-run actions, verified by a run of its own. If the user asks for both in one breath, finish this flow, report it, and then move on.

Declining the offer is a fine place to stop. Whether or not the user takes it, this flow's own job — build what was described, run it, report it, and persist or offer to persist it — is already done.

References

• references/preflight.md — establishing and announcing the session context.
• references/expectation-catalog.md — locating the shipped catalog, matching a described check against it, reading parameters, and what to do when nothing matches.
• references/robustness.md — the time-budget wrapper, scope-reduction levers, and how to report a failure helpfully.
• references/write-out.md — turning an in-memory session into a real project.