Skip down to the research record

Forecast refresh troubleshooting: 3 steps to trace totals

Quota Forecast Editorial Team · Published · 11 min read

To trace a changing forecast total, check 3 areas: the refresh scope, concurrent database reads, and HubSpot or Metabase retrieval and export limits. Microsoft Learn documents how to configure incremental refresh features in Power BI to capture fast-moving data and enhance performance. This sequence isolates where a changing total originates before you adjust any settings.

Key takeaways

  • Refresh scope: Check which partitions refresh on schedule. Microsoft Learn describes this incremental refresh configuration.
  • Concurrent reads: Check the isolation level your reporting connection uses. PostgreSQL documentation outlines these levels to govern how transactions view data changes from other concurrent transactions.
  • Retrieval limits: By default, search endpoints return pages of 10 records at a time. HubSpot specifies this page size for its CRM Search API endpoints.
  • Export row limit: By default, Metabase exports the first 1,048,575 rows of results. Metabase sets this boundary for exported data files.

First identify which Power BI partitions refresh

Start by distinguishing the import partitions that update on a schedule from any optional real-time query partition in your semantic model. Incremental refresh extends scheduled refresh operations by providing automated partition creation and management for semantic model tables that frequently load new and updated data, according to Microsoft Learn. This mechanism specifically targets tables where new rows or updates arrive regularly. When you see a total shift during the weekly call, check whether the change aligns with the boundary of these recent import partitions. The policy refreshes only the most recent import partitions, and optionally uses another DirectQuery partition for real-time data, which can significantly reduce the amount of data that has to be refreshed, as noted by Microsoft Learn. If your model includes this optional DirectQuery partition, it queries the source directly so that the latest changes at the data source appear in the query results, according to Microsoft Learn.

To trace the exact volume involved, look at the documented example of the FactInternetSales model table. With this type of policy, you can expect our FactInternetSales model table in the service, which averages 10,000 new rows per day, to refresh roughly 30,000 rows with each refresh operation, according to Microsoft Learn. This specific volume describes the behavior of that named model table under the described policy, not a universal standard for every firm's data warehouse. Do not assume your own model refreshes the same number of rows; instead, use this figure to understand how the policy limits the scope of the refresh to recent data points. The distinction between the scheduled import partitions and the optional real-time partition is critical because they handle data visibility differently. The import partitions contain historical and recently loaded data, while the DirectQuery partition pulls live data. If a forecast total changes unexpectedly, determine if the change falls within the window of the recent import partitions or if it originates from the live DirectQuery partition.

Then investigate changes between PostgreSQL reads

When a forecast total shifts between two checks, the database layer may be exposing data that was not stable at the first read. The PostgreSQL documentation defines specific transaction isolation behaviors that explain how concurrent operations affect what a query returns. You need to determine if your reporting query is reading uncommitted work or re-reading committed changes.

First, check for uncommitted data visibility. A transaction can read data written by a concurrent uncommitted transaction according to PostgreSQL. If your dashboard queries run while another process is writing to the same table but has not yet finished its transaction, the result set may include rows that do not persist. This creates a temporary total that disappears once the writing transaction rolls back or commits differently. Review the isolation level of your connection string to see if it permits this behavior.

Second, look for changed rereads within the same session. A transaction might re-read data it previously read and find that data has been modified by another transaction that committed since the initial read, as noted in PostgreSQL. If your application performs multiple queries in one session, such as fetching deal amounts and then recalculating totals, a commit from another user between those queries can alter the second result. This is distinct from the first scenario because the data was valid at the time of the first read but became stale before the second read.

To isolate this cause, compare the timestamp of the first query with the commit logs of other transactions active during that window. If you see writes committing between your read operations, the change is likely due to these isolation-level interactions rather than an extraction limit. Your next step is to verify whether your reporting tool runs each calculation in a single atomic transaction or splits them across separate connections.

Finally reconcile HubSpot retrieval and Metabase exports

When the total shifts between a CRM pull and a BI export, separate the API retrieval boundary from the file precision. The CRM Search API limits search endpoints to 10,000 total results for any given query HubSpot. If your pipeline contains more records than this cap, the API rejects paging beyond 10,000 results with a 400 error. A dashboard built on paged API calls may drop records without notice if its script ignores the 400 error. Check whether your extraction script handles this boundary or simply halts with an error code that goes unnoticed in logs.

On the reporting side, the exported file may contain data points that do not appear in the on-screen table. For example, if you formatted a floating point number to display only the first two decimal digits in the table results, exporting the unformatted results would include additional decimal digits (if any) found in the raw results Metabase. This discrepancy occurs because the visual layer applies formatting rules while the export function retrieves the underlying raw values. A forecast total calculated from displayed figures might differ slightly from one calculated from the CSV file due to these hidden decimal places. Verify whether your team sums the visible numbers or the exported rows during the weekly review.

If your dataset exceeds standard export capacities, the row count itself becomes a variable. For CSV exports, you can increase this limit with an environment variable: MB_DOWNLOAD_ROW_LIMIT Metabase. Without adjusting this setting, large datasets may truncate, causing the exported total to reflect only a subset of the pipeline. Compare the record count in your source system against the row count in the final CSV to identify truncation. For more checks on missing rows, see Forecast extract validation: 5 checks for missing rows. If the counts match but the totals differ, focus on the decimal precision difference described above. If the counts do not match, investigate the row limit configuration before attributing the variance to calculation errors.

To isolate the cause, run a controlled test: extract the same date range using both the API and the BI tool. Record the total record count from each source. Then sum the deal amounts using the displayed precision and again using the exported raw values. Note which method aligns with the official forecast figure. This process distinguishes between missing records due to API caps and value differences due to formatting. It also reveals if the environment variable needs adjustment for larger pipelines. Keep the resulting CSV and API log as evidence for the next call, so you can demonstrate exactly where the divergence occurred.

Product behavior reference table for changing totals

The following table lists the documented limits and behaviors for each publisher mentioned in the investigation. Each row cites the specific constraint or setting that can affect a forecast total during refresh or export operations.

PublisherDocumented Behavior or Limit
Microsoft LearnExample for the FactInternetSales model: loads two days of data (roughly 20,000 rows) initially; refreshes roughly 30,000 rows with each refresh operation Microsoft Learn.
PostgreSQLThe SQL standard defines four isolation levels; transactions may read uncommitted data or reread committed changes PostgreSQL.
HubSpotReturns pages of 10 records by default; limited to 10,000 total results for any given query HubSpot.
MetabaseExports first 1,048,575 rows by default; CSV limit adjustable through MB_DOWNLOAD_ROW_LIMIT Metabase.

The Microsoft Learn figures come from one named example model, not a baseline or default configuration. The Microsoft Learn example specifies initial load volumes and refresh operation sizes for its FactInternetSales model. The HubSpot API documentation distinguishes between the default page size and the maximum total result count allowed for any single query. The Metabase documentation identifies the default row boundary for exports and notes that exported files may contain more decimal precision than displayed in the interface. Review your own system settings against these documented limits to identify where your pipeline might diverge from expected behavior.

Keep the refresh, retrieval and precision evidence

Trace one changing forecast total through the reporting components your team uses, recording the refresh scope, retrieval boundary and displayed compared with exported precision. Start by writing down which partitions your model refreshes before the weekly call, then note the exact query limits applied during extraction. Record the retrieval boundary in your notes so you can distinguish a failed retrieval from a data change. If your team relies on Metabase for exports, check whether the CSV row limit has been adjusted through the named environment variable, as the default boundary is 1,048,575 rows. Note any difference between raw decimal precision in the file and the formatted numbers shown on screen. These two checks, retrieval errors and export boundaries, can explain some apparent shifts in totals. Keep this evidence alongside your refresh logs, and compare it with Reconcile sales dashboard totals with 4 checks when totals still disagree. When the next discrepancy appears, compare the recorded settings against the current state of each tool. If the refresh scope matches your notes but the retrieval boundary has changed, the issue likely lies in how data was pulled rather than how it was stored. This approach separates configuration drift from actual pipeline movement. Use the same tracking method for every weekly review to build a consistent history of your forecast components.

Related guides: Check 6 caption settings before the weekly sales call · Check 8 data rules before the weekly sales call · Verify 8 catalog steps for accurate forecast totals.

Investigating a changing forecast number: frequently asked questions

How does forecast refresh troubleshooting distinguish refresh scope from export limits?

Refresh scope defines which data partitions are updated during a scheduled operation, such as refreshing only the most recent import partitions Microsoft Learn. Export limits define the maximum number of rows a tool returns in a single file or query result. For example, Metabase exports the first 1,048,575 rows by default Metabase, while HubSpot search endpoints limit results to 10,000 total records for any given query HubSpot. Troubleshooting requires checking whether the change stems from the partitions that were refreshed or whether the retrieval boundary cut off part of the dataset.

Does my Power BI model use recent import partitions or an optional real-time query partition?

An incremental refresh policy can partition the table to refresh only the most recent import partitions Microsoft Learn. This policy may also include an optional DirectQuery partition for real-time data. The configuration determines whether the model relies on loaded historical data or queries the source directly for current changes. Check your semantic model settings to see if both partition types are active for the specific table involved in the forecast total.

What happens when a HubSpot search pages beyond 10,000 results?

Paging past that boundary returns a 400 error. By default, these endpoints return pages of 10 records at a time. If your forecast logic needs more than 10,000 records from a single search query, split the query into smaller searches.

Can a Metabase export contain more decimal places than the displayed forecast?

Yes, exporting unformatted results includes additional decimal digits found in the raw results Metabase. If you formatted a floating-point number to display only the first two decimal digits in the table view, the exported CSV file may contain more precision. This difference between displayed formatting and raw data can cause totals to appear different when compared across interfaces. Verify whether your forecast calculations use the rounded display values or the full-precision exported values.

What is a PostgreSQL transaction reread that encounters a committed change?

A transaction re-reads data it has previously read and finds that data has been modified by another transaction that committed since the initial read PostgreSQL. The SQL standard defines four levels of transaction isolation. A transaction might also read data written by a concurrent uncommitted transaction. Understanding which isolation level your database uses helps determine if such rereads affect the consistency of your forecast data during concurrent updates.

Sources