# Event Time vs Processing Time: Decide Which Date a Business Report Measures

<a id="main-content"></a>

Operations &amp; Systems

October 28, 2026  ·  7 min read  ·  Victor Valentine Romo

On this page

1. [Name the Five Times Behind One Observation](<#name-the-five-times-behind-one-observation>)
2. [State the Interval Before Counting Rows](<#state-the-interval-before-counting-rows>)
3. [Follow the Same Six Completions Across the Boundary](<#follow-the-same-six-completions-across-the-boundary>)
4. [Calculate Two Totals With Different Questions](<#calculate-two-totals-with-different-questions>)
5. [Use Beam Terminology Without Assuming a Product Contract](<#use-beam-terminology-without-assuming-a-product-contract>)
6. [Copy a Timestamp Policy Worksheet](<#copy-a-timestamp-policy-worksheet>)
7. [Review Missing and Conflicting Times Before Publication](<#review-missing-and-conflicting-times-before-publication>)
8. [Carry the Definition Into the Next Report](<#carry-the-definition-into-the-next-report>)
9. [Sources](<#sources>)

A service finishes before midnight. Its record reaches the reporting system the next morning. Two reports assign it to different months.

Both calculations can follow their instructions. The disagreement starts with the question each report answers.

Write the question before selecting the date column. A count of completed services and a count of received completion records measure different activity. Give each one a label that carries that distinction.

The definitions and worksheet below form a proposed reporting method. The six service records are fictional. Apache Beam supplies one documented implementation example. No particular reporting platform is required.

<a id="name-the-five-times-behind-one-observation"></a>

## Name the Five Times Behind One Observation

For this method, keep five timestamp meanings separate. A source might expose several, one, or none. These labels do not establish what [an undocumented field](</articles/csv-data-contract-business-handoffs/>) means.

| Timestamp | Meaning in this method | Evidence to retain |
| --- | --- | --- |
| Occurrence time | When the defined business event happened | Completion evidence and the field recording that event |
| Source-recorded time | When the source saved a record of the event | Source field definition and recorded value |
| Receipt time | When the reporting intake received the record | Intake timestamp and named receiving boundary |
| Processing time | When a named processing stage handled the record | Stage name, execution evidence, and timestamp |
| Report extraction time | When the report's input snapshot was extracted | Snapshot identifier and extraction timestamp |

A completed service can have a completion time distinct from its source-recorded time. In a fictional offline workflow, completion happens first. An operator saves the record after reconnecting. The receiving system then imports it.

Do not treat a generic `created_at` field as proof of occurrence time. Ask what creates the object and what action sets that field. Keep the answer unknown when the source definition is missing.

Name the receiving boundary too. Receipt by an integration gateway differs from receipt by a reporting dataset. Select one boundary for the metric. Preserve the other timestamp under its own name if it helps investigate delays.

Processing time needs a stage. A record can pass through intake validation, transformation, and report preparation. “Processed yesterday” leaves the relevant step unresolved. Write “validated at intake” when that is the event you mean.

Put these decisions in the [business metric dictionary](</articles/business-metric-dictionary-template/>). Keep the timestamp policy beside the metric's counting unit and eligibility rules.

<a id="state-the-interval-before-counting-rows"></a>

## State the Interval Before Counting Rows

Use a start-inclusive, end-exclusive interval for this example. September begins at 2026.09.01 00:00 UTC. It ends immediately before 2026.10.01 00:00 UTC.

An event exactly at the October boundary belongs to October. Do not assign it to September by changing the end to an imprecise “end of day.”

This example uses UTC for every displayed timestamp. A real owner can select another reporting zone. Record that zone explicitly before deriving the interval.

The scheduled preparation time is a separate decision. The [time-zone scheduling article](</articles/automation-time-zones-daylight-saving/>) covers when a job runs. Here, the question is which period receives each observation.

Record an availability cutoff as well. A September occurrence cannot contribute to an extracted report before that report has its record. Period assignment and available evidence answer separate questions.

For the calculations below, the snapshot includes receipts before 2026.10.03 08:00 UTC. Extraction occurs at 2026.10.03 09:00 UTC. Every listed record is available. No [later arrivals](</articles/late-arriving-data-report-revisions/>) enter this example.

<a id="follow-the-same-six-completions-across-the-boundary"></a>

## Follow the Same Six Completions Across the Boundary

Each row below represents one distinct completed service. There are no cancellations, reopened services, or duplicates. All records and timestamps are invented.

Read the dates as UTC. The occurrence column records the fictional service completion. The receipt column records arrival at the fictional reporting intake.

| Service | Occurrence time, UTC | Receipt time, UTC | Occurrence month | Receipt month |
| --- | --- | --- | --- | --- |
| S201 | 2026.09.29 16:00 | 2026.09.29 16:05 | September | September |
| S202 | 2026.09.30 23:50 | 2026.10.01 00:20 | September | October |
| S203 | 2026.09.30 23:59 | 2026.10.02 08:00 | September | October |
| S204 | 2026.10.01 00:00 | 2026.10.01 00:03 | October | October |
| S205 | 2026.10.01 10:00 | 2026.10.01 10:02 | October | October |
| S206 | 2026.09.30 15:00 | 2026.09.30 15:04 | September | September |

S202 and S203 cross the month boundary between occurrence and receipt. Their completions stay in September under an occurrence rule. Their receipts belong to October under a receipt rule.

S204 tests the boundary itself. Its occurrence is exactly 2026.10.01 00:00 UTC. The start-inclusive October interval includes it. September excludes it.

The table omits source-recorded and processing timestamps because neither calculation uses them. Do not invent those values to fill a diagram. Their meanings remain distinct even when this extract does not contain them.

Keep a copy of these six rows with both calculations. A reader should be able to reproduce the disagreement without comparing unrelated extracts.

<a id="calculate-two-totals-with-different-questions"></a>

## Calculate Two Totals With Different Questions

The first question is: “How many distinct services completed during September, using records available before the snapshot cutoff?”

Apply the occurrence-time rule. Include S201, S202, S203, and S206. The September completion count is four.

The second question is: “How many distinct completion records reached reporting intake during September?”

Apply the receipt-time rule to the same rows. Include S201 and S206. The September receipt count is two.

| Question | September members | September total | October members | October total |
| --- | --- | --- | --- | --- |
| When did the services complete? | S201, S202, S203, S206 | 4 | S204, S205 | 2 |
| When did reporting intake receive the records? | S201, S206 | 2 | S202, S203, S204, S205 | 4 |

October totals describe only the six-row example. They are not full-month forecasts or statements about all October activity.

Each rule assigns all six records exactly once across the displayed months. Four plus two equals six. Two plus four also equals six.

The September difference is two records. S202 and S203 explain it completely. No service has disappeared. The rule changed which month receives its contribution.

Use occurrence time for a question about when services completed. Use receipt time for a question about intake arrivals. The owner must select the business question before accepting either metric.

If two existing reports disagree, freeze their definitions and extracts first. Then use a [dashboard reconciliation bridge](</articles/reconcile-conflicting-business-dashboards/>) to isolate the records behind the difference.

<a id="use-beam-terminology-without-assuming-a-product-contract"></a>

## Use Beam Terminology Without Assuming a Product Contract

[Apache Beam's programming guide](<https://beam.apache.org/documentation/programming-guide/#watermarks-and-late-data>) distinguishes event time from processing time. Event time uses an element's timestamp; processing time follows the clock at the stage handling it.

That distinction helps explain these reporting choices. It does not prove that a business application's date field records occurrence. Verify the mapping before calling a field event time.

The service example compares occurrence and receipt. Receipt at a named intake is not automatically every later stage's processing time. Preserve that distinction in labels and documentation.

This article describes concepts from the online Beam guide checked on 2026.10.28. It specifies no SDK version or runner configuration. It makes no claim about a particular installed pipeline.

When discussing implementation, ask which field becomes the element timestamp. [A valid timestamp format](</articles/data-quality-dimensions-business-reports/>) does not establish a valid business meaning. Review the provenance with the person who owns the source event.

<a id="copy-a-timestamp-policy-worksheet"></a>

## Copy a Timestamp Policy Worksheet

Complete one worksheet per metric definition. Avoid a single company-wide “report date” field that hides different business questions.

| Policy field | Entry to complete | Filled synthetic example |
| --- | --- | --- |
| Metric and question | Stable identifier and plain question | `completed_services_v1`: when did services complete? |
| Counting unit | What contributes once | One distinct service identifier |
| Selected timestamp | Exact field and meaning | `completed_at`: documented service completion |
| Field provenance | Who sets it and from what evidence | Source completion action, retained source record |
| Reporting zone | Zone used to derive period boundaries | UTC |
| Boundary inclusivity | Start and end rules | Start included; end excluded |
| Availability cutoff | Which arrivals can enter this snapshot | Receipt before 2026.10.03 08:00 UTC |
| Extraction time | When this input copy was produced | 2026.10.03 09:00 UTC |
| Missing-field rule | Exclude, separate, or another explicit rule | Hold outside dated totals; count in an undated exception list |
| Conflicting-field rule | Who resolves incompatible evidence | Report owner reviews source completion evidence |
| Owner | Person accepting the meaning | Fictional service-report owner |
| Revision policy | What happens after new evidence arrives | Preserve original report and review a numbered revision |

Treat the filled column as a [worked example](</>). Its cutoff, ownership, and exception policy are editorial choices. Replace them with decisions appropriate to the actual report.

The missing-field rule deserves its own review. An undated completion remains undated. Substituting receipt time would change the question for that row.

If the owner needs an intake count, calculate it separately. Do not mix occurrence-assigned and receipt-assigned rows under a single completion label.

Preserve the original field value beside any interpreted value. [Record the transformation](</articles/data-lineage-small-business-reports/>) and its reason. A reviewer needs to distinguish supplied evidence from a reporting decision.

<a id="review-missing-and-conflicting-times-before-publication"></a>

## Review Missing and Conflicting Times Before Publication

Give the reviewer a short exception list with service identifiers and reasons. Keep private service details outside a broadly shared report.

For [a missing occurrence time](</articles/null-blank-zero-business-data/>), ask whether completion evidence exists elsewhere. Record the answer and its source. Leave the dated contribution unresolved until the policy permits inclusion.

For conflicting times, retain both values. A source correction may explain the difference. An undocumented edit may leave it unresolved. Do not select the value that makes a target look better.

If receipt appears earlier than occurrence, inspect the supplied zone, field definition, and source evidence. The ordering alone does not identify the cause. [Keep the anomaly visible](</articles/outlier-review-business-data/>) until someone can explain it.

Review exact-boundary records separately. Use S204 as the synthetic check: one timestamp at the next month's start, one month assigned. The report should not count it twice.

Finally, read the visible label without the worksheet. “September services” still hides the rule. “Services completed in September, available before the stated cutoff” tells the reader what the number measures.

<a id="carry-the-definition-into-the-next-report"></a>

## Carry the Definition Into the Next Report

Retain the timestamp policy version with each extracted report. A later policy change should name the old field, the new field, and the affected periods.

Keep changes in availability separate from changes in meaning. A new September completion record can revise an occurrence count without changing its definition. Switching from receipt to occurrence changes the metric definition itself.

For the six fictional rows, the definitions remain stable. September completion count is four. September intake receipt count is two. Each label points to its own question and reproducible membership.

Name the time the number measures.

<a id="sources"></a>

## Sources

- [Apache Beam: Programming Guide, watermarks and late data](<https://beam.apache.org/documentation/programming-guide/#watermarks-and-late-data>). Event-time and processing-time terminology. Online guide checked 2026.10.28. No SDK or runner configuration was tested.
