Redrock Postgres Documentation
Home Toggle Dark/Light/Auto mode Toggle Dark/Light/Auto mode Toggle Dark/Light/Auto mode Back to homepage

IO - DataFileRead

The DataFileRead event occurs when a connection waits on a backend process to read a required page from storage because the page isn’t available in shared memory.


All queries and data manipulation (DML) operations access pages in the buffer pool. Statements that can induce reads include SELECT, UPDATE, and DELETE. For example, an UPDATE can read pages from tables or indexes. If the page being requested or updated isn’t in the shared buffer pool, this read can lead to the DataFileRead event.

Because the shared buffer pool is finite, it can fill up. In this case, requests for pages that aren’t in memory force the database to read blocks from disk. If the DataFileRead event occurs frequently, your shared buffer pool might be too small to accommodate your workload. This problem is acute for SELECT queries that read a large number of rows that don’t fit in the buffer pool. For more information about the buffer pool, see shared buffer area.

Likely causes of increased waits

Common causes for the DataFileRead event include the following:

  • Connection spikes

    You might find multiple connections generating the same number of DataFileRead wait events. In this case, a spike (sudden and large increase) in DataFileRead events can occur.

  • SELECT and DML statements performing sequential scans

    Your application might be performing a new operation. Or an existing operation might change because of a new execution plan. In such cases, look for tables (particularly large tables) that have a greater seq_scan value. Find them by querying pg_stat_user_tables. To track queries that are generating more read operations, use the extension pg_stat_statements.

  • CTAS and CREATE INDEX for large data sets

    A CTAS is a CREATE TABLE AS SELECT statement. If you run a CTAS using a large data set as a source, or create an index on a large table, the DataFileRead event can occur. When you create an index, the database might need to read the entire object using a sequential scan. A CTAS generates DataFileRead events when pages aren’t in memory.

  • Ingesting large amounts of data

    When your application ingests large amounts of data, ANALYZE operations might occur more often. The ANALYZE process can be triggered by an autovacuum launcher or invoked manually.The ANALYZE operation reads a subset of the table. The number of pages that must be scanned is calculated by multiplying 30 by the default_statistics_target value. For more information, see the PostgreSQL documentation. The default_statistics_target parameter accepts values between 1 and 10,000, where the default is 100.

  • Resource starvation

    If instance network bandwidth or CPU are consumed, the DataFileRead event might occur more frequently.


We recommend different actions depending on the causes of your wait event.

Check predicate filters for queries that generate waits

Assume that you identify specific queries that are generating DataFileRead wait events. You might identify them using the following techniques:

  • Catalog views such as the one provided by the extension pg_stat_statements
  • The catalog view pg_stat_all_tables, if it periodically shows an increased number of physical reads
  • The pg_statio_all_tables view, if it shows that _read counters are increasing

We recommend that you determine which filters are used in the predicate (WHERE clause) of these queries. Follow these guidelines:

  • Run the EXPLAIN command. In the output, identify which types of scans are used. A sequential scan doesn’t necessarily indicate a problem. Queries that use sequential scans naturally produce more DataFileRead events when compared to queries that use filters.

    Find out whether the column listed in the WHERE clause is indexed. If not, consider creating an index for this column. This approach avoids the sequential scans and reduces the DataFileRead events. If a query has restrictive filters and still produces sequential scans, evaluate whether the proper indexes are being used.

  • Find out whether the query is accessing a very large table. In some cases, partitioning a table can improve performance, allowing the query to only read necessary partitions.

  • Examine the cardinality (total number of rows) from your join operations. Note how restrictive the values are that you’re passing in the filters for your WHERE clause. If possible, tune your query to reduce the number of rows that are passed in each step of the plan.

Minimize the effect of maintenance operations

Maintenance operations such as VACUUM and ANALYZE are important. We recommend that you don’t turn them off because you find DataFileRead wait events related to these maintenance operations. The following approaches can minimize the effect of these operations:

  • Run maintenance operations manually during off-peak hours. This technique prevents the database from reaching the threshold for automatic operations.
  • For very large tables, consider partitioning the table. This technique reduces the overhead of maintenance operations. The database only accesses the partitions that require maintenance.
  • When you ingest large amounts of data, consider disabling the autoanalyze feature.

The autoanalyze feature is automatically triggered for a table when the following formula is true.

pg_stat_user_tables.n_tup_ins > (pg_class.reltuples x autoanalyze_scale_factor) + autoanalyze_base_threshold

The view pg_stat_user_tables and catalog pg_class have multiple rows. One row can correspond to one row in your table. This formula assumes that the reltuples are for a specific table. The parameters autoanalyze_scale_factor (0.20 by default) and autoanalyze_base_threshold (50 tuples by default) are usually set globally for the whole instance. However, you can set different values for a specific table.

Respond to high numbers of connections

When you monitor database health status, you might find that the database connections metric spikes. This increase indicates an increased number of connections to your database. We recommend the following approach:

  • Limit the number of connections that the application can open with each instance. If your application has an embedded connection pool feature, set a reasonable number of connections. Base the number on what the CPUs in your instance can parallelize effectively.

  • Whenever possible, take advantage of hot standby nodes for PostgreSQL. When your application runs a read-only operation, send these requests to the reader-only endpoint. This technique spreads application requests across all reader nodes, reducing the I/O pressure on the writer node.

  • Consider scaling up your DB instance. A higher-capacity instance class gives more memory, which gives PostgreSQL a larger shared buffer pool to hold pages. The larger size also gives the DB instance more CPUs to handle connections. More CPUs are particularly helpful when the operations that are generating DataFileRead wait events are writes.