planetpostgresql Archives - credativ®


Tables that are created and dropped on demand, whether they are temporary or regular, are frequently used by application developers in PostgreSQL to simplify the implementation of various functionalities and to expedite responses. Numerous articles on the internet describe the advantages of using such tables for storing search results, precalculating figures for reports, importing data from external files, and more. One can even define a TEMP TABLE with the condition ON COMMIT DROP, allowing the system to clean up automatically. However, like most things, this solution has potential drawbacks, because size matters. A solution that functions smoothly for dozens of parallel sessions may suddenly begin to cause unexpected issues if the application is used by hundreds or thousands of users simultaneously during peak hours. Frequently creating and dropping tables and related objects, can cause significant bloat of certain PostgreSQL system tables. This is a well-known problem that many articles mention, but they often lack detailed explanations and quantification of the impact. Several pg_catalog system tables can become significantly bloated. Table pg_attribute is the most affected, followed by pg_attrdef and pg_class.

What is the main issue with the bloating of system tables?

We already encountered this issue in the PostgreSQL logs of one of our clients. When the bloat of system tables became too extensive, PostgreSQL decided to reclaim free space during an autovacuum operation. This action caused exclusive locks on the table and blocked all other operations for several seconds. PostgreSQL was unable to read information about the structures of all relations. And as a result, even the simplest select operations had to be delayed until the operation was resolved. This is, of course, an extreme and rare scenario that can only occur under exceptionally high load. Nevertheless, it’s important to be aware of it and be able to assess if it could also happen to our database.

Example of reporting table in accounting software

Let’s examine the impact of these short-lived relations on PostgreSQL system tables using two different examples. The first is a comprehensive example of TEMP TABLE where we will explain all the details, and the second is for benchmarking purposes. Our first example involves an imaginary accounting software that generates a wide variety of reports, many of which require some precalculation of results. The use of temporary tables for these purposes is a fairly obvious design choice. We will discuss one such example — a temporary pivot table for a report storing monthly summaries for an entire year, with one row per client_id:
CREATE TEMP TABLE pivot_temp_table (
   id serial PRIMARY KEY,
   inserted_at timestamp DEFAULT current_timestamp,
   client_id INTEGER,
   name text NOT NULL,
   address text NOT NULL,
   loyalty_program BOOLEAN DEFAULT false,
   loyalty_program_start TIMESTAMP,
   orders_202301_count_of_orders INTEGER DEFAULT 0,
   orders_202301_total_price NUMERIC DEFAULT 0,
   orders_202312_count_of_orders INTEGER DEFAULT 0,
   orders_202312_total_price NUMERIC DEFAULT 0);
We also want to create some indexes because some results can be quite huge:
CREATE INDEX pivot_temp_table_idx1 ON pivot_temp_table (client_id);
CREATE INDEX pivot_temp_table_idx2 ON pivot_temp_table (name);
CREATE INDEX pivot_temp_table_idx3 ON pivot_temp_table (loyalty_program);
CREATE INDEX pivot_temp_table_idx4 ON pivot_temp_table (loyalty_program_start);
Summary of the created objects:
  • A temporary table, pivot_temp_table, with 31 columns, 27 of which have default values.
  • Some of the columns are of the TEXT data type, resulting in the automatic creation of a TOAST table.
  • The TOAST table requires an index on chunk_id and chunk_seq.
  • The ID is the primary key, meaning a unique index on ID was automatically created.
  • The ID is defined as SERIAL, leading to the automatic creation of a sequence, which is essentially another table with a special structure.
  • We also defined four additional indexes on our temporary table.

Let’s now examine how these relations are represented in PostgreSQL system tables.

Table pg_attribute

The pg_attribute table stores the attributes (columns) of all relations. PostgreSQL will insert a total of 62 rows into the pg_attribute table:
  • Each row in our pivot_temp_table contains six hidden columns (tableoid, cmax, xmax, cmin, xmin, ctid) and 31 ‘normal’ column. This totals to 37 rows inserted for the main temp table.
  • Indexes will add one row for each column used in the index, equating to five rows in our case.
  • A TOAST table was automatically created. It has six hidden columns and three normal columns (chunk_id, chunk_seq, chunk_data), and one index on chunk_id and chunk_seq, adding up to 11 rows in total.
  • A sequence for the ID was created, which is essentially another table with a predefined structure. It has six hidden columns and three normal columns (last_value, log_cnt, is_called), adding another nine rows.

Table pg_attrdef

The pg_attrdef table stores default values for columns. Our main table contains many default values, resulting in the creation of 27 rows in this table. We can examine their content using a query:
   c.relname as table_name,
   o.rolname as table_owner,
   c.relkind as table_type,
   a.attname as column_name,
   a.attnum as column_number,
   a.atttypid::regtype as column_data_type,
   pg_get_expr(adbin, adrelid) as sql_command
FROM pg_attrdef ad
JOIN pg_attribute a ON ad.adrelid = a.attrelid AND ad.adnum = a.attnum
JOIN pg_class c ON c.oid = ad.adrelid
JOIN pg_authid o ON o.oid = c.relowner
WHERE c.relname = 'pivot_temp_table'
ORDER BY table_name, column_number;
Our output:
    table_name    | table_owner | table_type |         column_name           | column_number |     column_data_type        | sql_command
 pivot_temp_table | postgres    | r          | id                            | 1             | integer                     | nextval('pivot_temp_table_id_seq'::regclass)
 pivot_temp_table | postgres    | r          | inserted_at                   | 2             | timestamp without time zone | CURRENT_TIMESTAMP
 pivot_temp_table | postgres    | r          | loyalty_program               | 6             | boolean                     | false
 pivot_temp_table | postgres    | r          | orders_202301_count_of_orders | 8             | integer                     | 0
 pivot_temp_table | postgres    | r          | orders_202301_total_price     | 9             | numeric                     | 0
--> up to the column "orders_202312_total_price"

Table pg_class

The pg_class table stores primary information about relations. This example will create nine rows: one for the temp table, one for the toast table, one for the toast table index, one for the ID primary key unique index, one for the sequence, and four for the custom indexes.

Summary of this example

Our first example produced a seemingly small number of rows – 62 in pg_attribute, 27 in pg_attrdef, and 9 in pg_class. These are very low numbers, and if such a solution was used by only one company, we would hardly see any problems. But consider a scenario where a company hosts accounting software for small businesses and hundreds or even thousands of users use the app during peak hours. In such a situation, many temp tables and related objects would be created and dropped at a relatively quick pace. In the pg_attribute table, we could see anywhere from a few thousand to even hundreds of thousands of records inserted and then deleted over several hours. However, this is still a relatively small use case. Let’s now imagine and benchmark something even larger.

Example of online shop

Let’s conduct deeper analysis using a more relatable and heavier example. Imagine an online retailer selling clothing, shoes, and other accessories. When a user logs into the shop, the database automatically creates some user-specific tables. These are later deleted by a dedicated process after a certain period of user inactivity. These relations are created to speed up the system’s responses to a specific user. Repeated selects from the main tables would be much slower, even though the main tables are partitioned by days, these partitions can be enormous. For our example, we don’t need to discuss the layout of sessions, nor whether the tables are created as temporary or regular ones, as both have the same impact on PostgreSQL system tables. We will also omit all other aspects of real-life implementation. This example is purely theoretical, inspired by design patterns discussed on the internet, and is not based on any real system. It should not be understood as a design recommendation. In fact, as we will see, this example would more likely serve as an anti-pattern.
  1. The “session_events” table stores selected actions performed by the user during the session. Events are collected for each action the user takes on the website, so there are at least hundreds, but more often thousands of events recorded from one session. These are all sent in parallel into the main event table. However, the main table is enormous. Therefore, this user-specific table stores only some events, allowing for quick analysis of recent activities, etc. The table has 25 different columns, some of which are of the TEXT type and one column of the JSONB type – which means a TOAST table with one index was created. The table has a primary key of the serial type, indicating the order of actions – i.e., one unique index, one sequence, and one default value were created. There are no additional default values. The table also has three additional indexes for quicker access, each on one column. Their benefit could be questionable, but they are part of the implementation.
    • Summary of rows in system tables – pg_attribute – 55 rows, pg_class – 8 rows, pg_attrdef – 1 row
  2. The “last_visited” table stores a small subset of events from the “session_events” table to quickly show which articles the user has visited during this session. Developers chose to implement it this way for convenience. The table is small, containing only 10 columns, but at least one is of the TEXT type. Therefore, a TOAST table with one index was created. The table has a primary key of the TIMESTAMP type, therefore it has one unique index, one default value, but no sequence. There are no additional indexes.
    • Rows in system tables – pg_attribute – 28 rows, pg_class – 4 rows, pg_attrdef – 1 row
  3. The “last_purchases” table is populated at login from the main table that stores all purchases. This user-specific table contains the last 50 items purchased by the user in previous sessions and is used by the recommendation algorithm. This table contains fully denormalized data to simplify their processing and visualization, and therefore it has 35 columns. Many of these columns are of the TEXT type, so a TOAST table with one index was created. The primary key of this table is a combination of the purchase timestamp and the ordinal number of the item in the order, leading to the creation of one unique index but no default values or sequences. Over time, the developer created four indexes on this table for different sorting purposes, each on one column. The value of these indexes can be questioned, but they still exist.
    • Rows in system tables – pg_attribute – 57 rows, pg_class – 8 rows
  4. The “selected_but_not_purchased” table is populated at login from the corresponding main table. It displays the last 50 items still available in the shop that the user previously considered purchasing but later removed from the cart or didn’t finish ordering at all, and the content of the cart expired. This table is used by the recommendation algorithm and has proven to be a successful addition to the marketing strategy, increasing purchases by a certain percentage. The table has the same structure and related objects as “last_purchases”. Data are stored separately from purchases to avoid mistakes in data interpretation and also because this part of the algorithm was implemented much later.
    • Rows in system tables – pg_attribute – 57 rows, pg_class – 8 rows
  5. The “cart_items” table stores items selected for purchase in the current session but not yet bought. This table is synchronized with the main table, but a local copy in the session is also maintained. The table contains normalized data, therefore it has only 15 columns, some of which are of the TEXT type, leading to the creation of a TOAST table with one index. It has a primary key ID of the UUID type to avoid collisions across all users, resulting in the creation of one unique index and one default value, but no sequence. There are no additional indexes.
    • Rows in system tables – pg_attribute – 33 rows, pg_class – 4 rows, pg_attrdef – 1 row

The creation of all these user-specific tables results in the insertion of the following number of rows into PostgreSQL system tables – pg_attribute: 173 rows, pg_class: 32 rows, pg_attrdef: 3 rows.

Analysis of traffic

As the first step we provide an analysis of the business use case and traffic seasonality. Let’s imagine our retailer is active in several EU countries and targets mainly people from 15 to 35 years old. The online shop is relatively new, so it currently has 100,000 accounts. Based on white papers available on the internet, we can presume the following user activity:

Level of user’s activity Ratio of users [%] Total count of users Frequency of visits on page
very active 10% 10,000 2x to 4x per week
normal activity 30% 30,000 ~1 time per week
low activity 40% 40,000 1x to 2x per month
almost no activity 20% 20,000 few times in year

Since this is an online shop, traffic is highly seasonal. Items are primarily purchased by individuals for personal use. Therefore, during the working day, they check the shop at very specific moments, such as during travel or lunchtime. The main peak in traffic during the working day is between 7pm and 9pm. Fridays usually have much lower traffic, and the weekend follows suit. The busiest days are generally at the end of the month, when people receive their salaries. The shop experiences the heaviest traffic on Thanksgiving Thursday and Black Friday. The usual practice in recent years is to close the shop for an hour or two and then reopen at a specific hour with reduced prices. Which translates into huge number of relations being created and later deleted at relatively short time. The duration of a user’s connection can range from just a few minutes up to half an hour. User-specific tables are created when user logs into shop. They are later deleted by a special process that uses a sophisticated algorithm to determine whether relations already expired or not. This process involves various criteria and runs at distinct intervals, so we can see a large number of relations deleted in one run. Let’s quantify these descriptions:

Traffic on different days Logins per 30 min pg_attribute [rows] pg_class [rows] pg_attrdef [rows]
Numbers from analysis per 1 user 1 173 32 3
Average traffic in the afternoon 1,000 173,000 32,000 3,000
Normal working day evening top traffic 3,000 519,000 96,000 9,000
Evening after salary low traffic 8,000 1,384,000 256,000 24,000
Evening after salary high traffic 15,000 2,595,000 480,000 45,000
Singles’ Day evening opening 40,000 6,920,000 1,280,000 120,000
Thanksgiving Thursday evening opening 60,000 10,380,000 1,920,000 180,000
Black Friday evening opening 50,000 8,650,000 1,600,000 150,000
Black Friday weekend highest traffic 20,000 3,460,000 640,000 60,000
Theoretical maximum – all users connected 100,000 17,300,000 3,200,000 300,000

Now we can see what scalability means. Our solution will definitely work reasonably on normal days. However, traffic in the evenings after people receive their salaries can be very heavy. Thanksgiving Thursday and Black Friday really test its limits. Between 1 and 2 million user-specific tables and related objects will be created and deleted during these evenings. And what happens if our shop becomes even more successful and the number of accounts grows to 500 000, 1 million or more? The solution would definitely hit the limits of vertical scaling at some points, and we would need to think about ways to scale it horizontally.

How to examine bloat

Analysis of traffic provided some theoretical numbers. But we need to check the real-time situation in our database. First, if we’re unsure about what’s happening in our system regarding the creation and deletion of relations, we can temporarily switch on extended logging. We can set ‘log_statements’ to at least ‘ddl’ to see all CREATE/ ALTER /DROP commands. To monitor long running vacuum actions we can set ‘log_autovacuum_min_duration’ to some reasonable low number like 2 seconds. These settings are both dynamic and do not require a restart. However, this change may increase disk IO on local servers due to the increased writes into PostgreSQL logs. On cloud databases or Kubernetes clusters, log messages are usually sent to a separate subsystem and stored independently of the database disk, so the impact should be minimal. To check existing bloats in PostgreSQL tables, we can use the ‘pgstattuple’ extension. This extension only creates new functions; it does not influence the performance of the database. It can only cause reads when we invoke some of its functions. By using its functions in combination with results from other PostgreSQL system objects, we can get a better picture of the bloat in the PostgreSQL system tables. The pg_relation_size function was added to double-check the numbers from the pgstattuple function.
WITH tablenames AS (SELECT tablename FROM (VALUES('pg_attribute'),('pg_attrdef'),('pg_class')) as t(tablename))
   now() as checked_at,
   pg_relation_size(tablename) as relation_size,
   pg_relation_size(tablename) / (8*1024) as relation_pages,
FROM tablenames t
JOIN LATERAL (SELECT * FROM pgstattuple(t.tablename)) s ON true
JOIN LATERAL (SELECT last_autovacuum, last_vacuum, last_autoanalyze, last_analyze, n_live_tup, n_dead_tup
FROM pg_stat_all_tables WHERE relname = t.tablename) a ON true
ORDER BY tablename
We will get output like this one (result is shown only for 1 table)
 tablename         | pg_attribute
 checked_at        | 2024-02-18 10:46:34.348105+00
 relation_size     | 44949504
 relation_pages    | 5487
 last_autovacuum   | 2024-02-16 20:07:15.7767+00
 last_vacuum       | 2024-02-16 20:55:50.685706+00
 last_autoanalyze  | 2024-02-16 20:07:15.798466+00
 last_analyze      | 2024-02-17 22:05:43.19133+00
 n_live_tup        | 3401
 n_dead_tup        | 188221
 table_len         | 44949504
 tuple_count       | 3401
 tuple_len         | 476732
 tuple_percent     | 1.06
 dead_tuple_count  | 107576
 dead_tuple_len    | 15060640
 dead_tuple_percent| 33.51
 free_space        | 28038420
 free_percent      | 62.38
If we attempt some calculations, we’ll find that the summary of numbers from the pgstattuple function does not match the total relation size. Also, the percentages usually don’t add up to 100%. We need to understand these values as estimates, but they still provide a good indication of the scope of the bloat. We can easily modify this query for monitoring purposes. We should certainly monitor at least the relation_size, n_live_tup, and n_dead_tup for these system tables. To run monitoring under a non-superuser account, this account must have been granted or inherited PostgreSQL predefined roles ‘pg_stat_scan_tables’ or ‘pg_monitor’. If we want to dig deeper into the problem and make some predictions, we can, for example, check how many tuples are stored per page in a specific table. With these numbers, we would be able to estimate possible bloat in critical moments. We can use a query like this one:
WITH pages AS (
   SELECT * FROM generate_series(0, (SELECT pg_relation_size('pg_attribute') / 8192) -1) as pagenum),
tuples_per_page AS (
   SELECT pagenum, nullif(sum((t_xmin is not null)::int), 0) as tuples_per_page
   FROM pages JOIN LATERAL (SELECT * FROM heap_page_items(get_raw_page('pg_attribute',pagenum))) a ON true
   GROUP BY pagenum)
   count(*) as pages_total,
   min(tuples_per_page) as min_tuples_per_page,
   max(tuples_per_page) as max_tuples_per_page,
   round(avg(tuples_per_page),0) as avg_tuples_per_page,
   mode() within group (order by tuples_per_page) as mode_tuples_per_page
FROM tuples_per_page
Output will look like this:
 pages_total          | 5487
 min_tuples_per_page  | 1
 max_tuples_per_page  | 55
 avg_tuples_per_page  | 23
 mode_tuples_per_page | 28

Here, we can see that in our pg_attribute system table, we have an average of 23 tuples per page. So now we can calculate theoretical increase in size of this table for different traffic. Typical size of this table is usually only few hundreds of MBs. So theoretical bloat about 3 GB during Black Friday days is quite significant number for this table.

Logins pg_attribute rows data pages size in MB
1 173 8 0.06
1,000 173,000 7,522 58.77
3,000 519,000 22,566 176.30
15,000 2,595,000 112,827 881.46
20,000 3,460,000 150,435 1,175.27
60,000 10,380,000 451,305 3,525.82
100,000 17,300,000 752,174 5,876.36


We’ve presented a reporting example from accounting software and an example of user-specific tables from an online shop. While both are theoretical, the idea is to illustrate patterns. We also discussed the influence of high traffic seasonality on the number of inserts and deletes in system tables. We provided an example of an extremely increased load in an online shop on big sales days. We believe the results of the analysis warrant attention. It’s also important to remember that the already heavy situation in these peak moments can be even more challenging if our application is running on an instance with low disk IOPS. All these new objects would cause writes into WAL logs and synchronization to the disk. In the case of low disk throughput, there could be significant latency, and many operations could be substantially delayed. So, what’s the takeaway from this story? First of all, PostgreSQL autovacuum processes are designed to minimize the impact on the system. If the autovacuum settings on our database are well-tuned, in most cases, we won’t see any problems. However, if these settings are outdated, tailored for much lower traffic, and our system is under unusually heavy load for an extended period, creating and dropping thousands of tables and related objects in a relatively short time, PostgreSQL system tables can eventually become significantly bloated. This will already slow down system queries reading details about all other relations. And at some point, the system could decide to shrink these system tables, causing an Exclusive lock on some of these relations for seconds or even dozens of seconds. This could block a large number of selects and other operations on all tables. Based on analysis of traffic, we can conduct a similar analysis for other specific systems to understand when they will be most susceptible to such incidents. But having effective monitoring is absolutely essential.


  1. Understanding an outage: concurrency control & vacuuming in PostgreSQL
  2. Stackoverflow – temporary tables bloating pg_attribute
  3. Diagnosing table and index bloat
  4. What are the peak times for online shopping?
  5. PostgreSQL Tuple-Level Statistics With pgstattuple

The PostgreSQL 2024Q1 back-branch releases 16.2, 15.6, 14.11, 13.14 and 12.18 on February 8th 2024. Besides fixing a security issue (CVE-2024-0985) and the usual bugs, they are somewhat unique in that they address two performance problems by backporting fixes already introduced into the master branch before. In this blog post, we describe two quick benchmarks that show how the new point releases have improved. The benchmarks were done on a ThinkPad T14s Gen 3 which has a Intel i7-1280P CPU with 20 cores and 32 GB of RAM.

Scalability Improvements During Heavy Contention

The performance improvements in the 2024Q1 point releases concerns locking scalability improvements at high client counts, i.e., when the system is under heavy contention. Benchmarks had shown that the performance was getting worse dramatically for a pgbench run with more than 128 clients. The original commit to master (which subsequently was released with version 16) is from November 2022. It got introduced into the back-branches now as version 16 has seen some testing and the results were promising.

The benchmark we used is adapted from this post by the patch author and consists of a tight pgbench run simply executing SELECT txid_current() for five seconds each at increasing client count and measuring the transactions per second:

$ cat /tmp/txid.sql
SELECT txid_current();
$ for c in 1 2 4 8 16 32 64 96 128 192 256 384 512 768 1024 1536;
> do echo -n "$c ";pgbench -n -M prepared -f /tmp/txid.sql -c$c -j$c -T5 2>&1|grep '^tps'|awk '{print $3}';
> done

The following graph shows the average transactions per second (tps) over 3 runs with increasing client count (1-1536 clients), using the Debian 12 packages for version 15, comparing the 2023Q4 release (15.5, package postgresql-15_15.5-0+deb12u1) with the 2024Q1 release (15.6, package postgresql-15_15.6-0+deb12u1):

The tps numbers are basically the same up to 128 clients, whereas afterwards the 15.5 transaction counts drops from the peak of 650k 10-fold to 65k. The new 15.6 release maintains the transaction count much better and still maintains around 300k tps at the 1536 clients, which is a 4.5-fold increase of the 2024Q1 release compared to previously.

This benchmark is of course a best-case, artificial scenario, but it shows that the latest point release of Postgres can improve scalability dramatically for heavily contested locking scenarios.

JIT Memory Consumption Improvements

JIT (just-in-time compilation with LLVM) was introduced in version 11 of Postgres and made the default in version 13. For a long time now, it has been known that long-running PostgreSQL sessions that run JIT queries repeatedly leak memory. There have been several bug reports about this, including some more in the Debian bug tracker and probably elsewhere.

This has been diagnosed to be due to JIT inlining and a work-around is setting jit_inline_above_cost to -1 from the default value of 500000. However, this disables JIT inlining completely. The 2024Q1 back-branch releases contain a backport of a change that will go into version 17: after every 100 queries, the LLVM caches are dropped and recreated, plugging the memory leak.

To show how the memory consumption has improved, we use the test case from this bug report. The benchmark is prepared as followed:

   id integer NOT NULL,
   CONSTRAINT leak_test_pkey PRIMARY KEY (id)

INSERT INTO leak_test(id)
   SELECT id
   FROM generate_series(1,100000) id

Then, the process ID of the backend is noted and the SQL query mentioned in the bug report run 5000 times in a loop:

=> SELECT pg_backend_pid();

=> DO $$DECLARE loop_cnt integer;
->   loop_cnt := 5000;
->   LOOP
->     PERFORM
->       id,
->       (SELECT count(*) FROM leak_test x WHERE as x_result,
->       (SELECT count(*) FROM leak_test y WHERE as y_result
->       /* Leaks memory around 80 kB on each query, but only if two sub-queries are used. */
->     FROM leak_test l;
->     loop_cnt := loop_cnt - 1;
->     EXIT WHEN loop_cnt = 0;
->   END LOOP;
-> END$$;

During this the memory consumption of the Postgres backend is recorded via pidstat:

pidstat -r -hl -p 623404 2 | tee -a leak_test.log.15.6
Linux 6.1.0-18-amd64 (     15.02.2024  _x86_64_    (20 CPU)

# Time        UID       PID  minflt/s  majflt/s     VSZ     RSS   %MEM  Command
12:48:56      118    623404      0,00      0,00  381856   91504   0,28  postgres: 15/main: postgres postgres [local] SELECT
12:48:58      118    623404      0,00      0,00  381856   91504   0,28  postgres: 15/main: postgres postgres [local] SELECT
12:49:00      118    623404      0,00      0,00  381856   91504   0,28  postgres: 15/main: postgres postgres [local] SELECT
12:49:02      118    623404      0,00      0,00  381856   91504   0,28  postgres: 15/main: postgres postgres [local] SELECT
12:49:04      118    623404   7113,00      0,00  393632  109252   0,34  postgres: 15/main: postgres postgres [local] SELECT
12:49:06      118    623404  13219,00      0,00  394556  109508   0,34  postgres: 15/main: postgres postgres [local] SELECT
12:49:08      118    623404  14376,00      0,00  395384  108228   0,33  postgres: 15/main: postgres postgres [local] SELECT

The benchmark are again repeated for the 15.5 and 15.6 Debian 12 packages (which are both linked against LLVM-14) and the RSS memory consumption as reported by pidstat is plotted against time:

While the memory consumption of the 15.5 session increases linearly over time from 100 to 600 MB, it stays more or less constant at around 100 MB for 15.6. This is a great improvement that will make JIT much more usable for larger installations with long running sessions where so far the usual recommendation has been to disable JIT entirely.


The 2024Q1 patch release has important performance improvements for lock scalability and JIT memory consumption that we have demonstrated in this blog post. Furthermore, the patch release contains other important bug fixes and a security fix for CVE-2024-0985. This security issue is limited to materialized views and a admin user needs to be tricked into recreating a malicious materialized view on behalf of an attacker. But it has seen some german press coverage so quite a few of our customers were especially made aware of it and asked us to assist them with their minor upgrades. In general, Postgres patch releases are low-risk and unintrusive (just install the updated packages and restart the Postgres instances if the package did not do this itself) so that they should always be deployed as soon as possible.

Moodle is a popular Open Source online learning platform. Especially since the beginning of the COVID-19 pandemic the importance of Moodle for schools and universities has further increased. In some states in Germany all schools had to switch to Moodle and other platforms like BigBlueButton in the course of a few days. This leads to scalability problems if suddenly several tens of thousands of pupils need to access Moodle.
Besides scaling the Moodle application itself, the database needs to be considered as well. One of the database options for Moodle is PostgreSQL. In this blog post, we present load-balancing options for Moodle using PostgreSQL.

High-Availability via Patroni

An online learning platform can be considered critical infrastructure from the point of view of the educational system and should be made highly available, in particular the database. A good solution for PostgreSQL is Patroni, we reported on its Debian-integration in the past.

In short, Patroni uses a distributed consensus store (DCS) to elect a leader from a typically 3-node cluster or initiate a failover and elect a new leader in the case of a leader failure, without entering a split-brain scenario. In addition, Patroni provides a REST API used for communication among nodes and from the patronictl program, e.g. to change the Postgres configuration online on all nodes or to initiate a switchover.

Client-solutions for high availability

From Moodle’s perspective, however, it must additionally be ensured that it is connected to the leader, otherwise no write transactions are possible. Traditional high-availability solutions such as Pacemaker use virtual IPs (VIPs) here, which are pivoted to the new primary node in the event of a failover. For Patroni there is the vip-manager project instead, which monitors the leader key in the DCS and sets or removes cluster VIP locally. This is also integrated into Debian as well.

An alternative is to use client-side failover based on PostgreSQL’s libpq library. For this, all cluster members are listed in the connection string and the connection option target_session_attrs=read-write is added. Configured this way, if a connection is broken, the client will try to reach the other nodes until a new primary is found.

Another option is HAProxy, a highly scalable TCP/HTTP load balancer. By performing periodic health checks on Patroni‘s REST API of each node, it can determine the current leader and forward client queries to it.

Moodle database configuration

Moodle’s connection to a PostgreSQL database is configured in config.php, e.g. for a simple stand-alone database:

$CFG->dbtype    = 'pgsql';
$CFG->dblibrary = 'native';
$CFG->dbhost    = '';
$CFG->dbname    = 'moodle';
$CFG->dbuser    = 'moodle';
$CFG->dbpass    = 'moodle';
$CFG->prefix    = 'mdl_';
$CFG->dboptions = array (
  'dbport' => '',
  'dbsocket' => ''

The default port 5432 is used here.

If streaming replication is used, the standbys can additionally be defined as readonly and assigned to an own database user (which only needs read permissions):

$CFG->dboptions = array (
  'readonly' => [
    'instance' => [
      'dbhost' => '',
      'dbport' =>  '',
      'dbuser' => 'moodle_safereads',
      'dbpass' => 'moodle'
      'dbhost' => '',
      'dbport' =>  '',
      'dbuser' => 'moodle_safereads',
      'dbpass' => 'moodle'

Failover/load balancing with libpq

If a highly available Postgres cluster is used with Patroni, the primary, as described above, can be switched to prevent loss of data or shutdown of the system, in case of a failover or switchover incident. Moodle does not provide a way to set generic database options here and thus setting target_session_attrs=read-write directly is not possible. Therefore we developed a patch for this and implemented it in the Moodle tracker. This allows the additional option 'dbfailover' => 1, in the $CFG->dboptions array, which adds the necessary connection option target_session_attrs=read-write. A customized config.php would look like this:

$CFG->dbtype    = 'pgsql';
$CFG->dblibrary = 'native';
$CFG->dbhost    = ',,';
$CFG->dbname    = 'moodle';
$CFG->dbuser    = 'moodle';
$CFG->dbpass    = 'moodle';
$CFG->prefix    = 'mdl_';
$CFG->dboptions = array (
  'dbfailover' => 1,
  'dbport' => '',
  'dbsocket' => '',
  'readonly' => [
    'instance' => [
      'dbhost' => '',
      'dbport' =>  '',
      'dbuser' => 'moodle_safereads',
      'dbpass' => 'moodle'
      'dbhost' => '',
      'dbport' =>  '',
      'dbuser' => 'moodle_safereads',
      'dbpass' => 'moodle'
      'dbhost' => '',
      'dbport' =>  '',
      'dbuser' => 'moodle_safereads',
      'dbpass' => 'moodle'

Failover/load balancing with HAProxy

If HAProxy is to be used instead, then $CFG->dbhost must be set to the HAProxy host e.g. in case HAProxy is running locally on the Moodle server(s). Moreover a second port (e.g. 65432) can be defined for read queries, which is configured as readonly in $CFG->dboptions, same as the streaming replication standby above. The config.php would then look like this:

$CFG->dbtype    = 'pgsql';
$CFG->dblibrary = 'native';
$CFG->dbhost    = '';
$CFG->dbname    = 'moodle';
$CFG->dbuser    = 'moodle';
$CFG->dbpass    = 'moodle';
$CFG->prefix    = 'mdl_';
$CFG->dboptions = array (
  'dbport' => '',
  'dbsocket' => '',
  'readonly' => [
    'instance' => [
      'dbhost' => '',
      'dbport' =>  '65432',
      'dbuser' => 'moodle_safereads',
      'dbpass' => 'moodle'

The HAProxy configuration file haproxy.cfg can look like the following example:

    maxconn 100

    log global
    mode tcp
    retries 2
    timeout client 30m
    timeout connect 4s
    timeout server 30m
    timeout check 5s

listen stats
    mode http
    bind *:7000
    stats enable
    stats uri /

listen postgres_write
    bind *:5432
    mode tcp
    option httpchk
    http-check expect status 200
    default-server inter 3s fall 3 rise 3 on-marked-down shutdown-sessions
    server pg1 maxconn 100 check port 8008
    server pg2 maxconn 100 check port 8008
    server pg3 maxconn 100 check port 8008

HAProxy expects incoming write connections (postgres_write) on port 5432 and forwards them to port 5432 of the cluster members. The primary is determined by an HTTP check on port 8008 (the default Patroni REST API port); Patroni returns status 200 here for the primary and status 503 for standbys.

For read queries (postgres_read), it must be decided whether the primary should also serve read-only queries or not. If this is the case, a simple Postgres check (pgsql-check) can be used; however, this may lead to entries in the PostgreSQL log regarding incorrect or incomplete logins:

listen postgres_read
    bind *:65432
    mode tcp
    balance leastconn
    option pgsql-check user haproxy
    default-server inter 3s fall 3 rise 3 on-marked-down shutdown-sessions
    server pg1 check
    server pg2 check
    server pg3 check

If you don’t want the primary to participate in the read scaling you can simply use the same HTTP check as in the postgres_write section, this time expecting HTTP status 503:

listen postgres_read
    bind *:65432
    mode tcp
    balance leastconn
    option httpchk
    http-check expect status 503
    default-server inter 3s fall 3 rise 3 on-marked-down shutdown-sessions
    server pg1 check port 8008
    server pg2 check port 8008
    server pg3 check port 8008

Revised Ansible playbook

HAProxy support has also been implemented in version 0.3 of our Ansible playbooks for automated setup of a three-node PostgreSQL Patroni cluster on Debian. The new variable haproxy_primary_read_scale can be used to decide whether HAProxy should also issue requests on the read-only port to the primary node or only to the followers.

We are happy to help!

Whether it’s PostgreSQL, Patroni, HAProxy, Moodle, or any other open source software; with over 22+ years of development and service experience in the open source space, credativ GmbH can assist you with unparalleled and individually customizable support. We are there to help and assist you in all your open source infrastructure needs – if desired 24 hours a day, 365 days a year!

We look forward to hearing from you.

SQLreduce: Reduce verbose SQL queries to minimal examples

Developers often face very large SQL queries that raise some errors. SQLreduce is a tool to reduce that complexity to a minimal query.

SQLsmith generates random SQL queries

SQLsmith is a tool that generates random SQL queries and runs them against a PostgreSQL server (and other DBMS types). The idea is that by fuzz-testing the query parser and executor, corner-case bugs can be found that would otherwise go unnoticed in manual testing or with the fixed set of test cases in PostgreSQL’s regression test suite. It has proven to be an effective tool with over 100 bugs found in different areas in the PostgreSQL server and other products since 2015, including security bugs, ranging from executor bugs to segfaults in type and index method implementations. For example, in 2018, SQLsmith found that the following query triggered a segfault in PostgreSQL:

  case when pg_catalog.lastval() < pg_catalog.pg_stat_get_bgwriter_maxwritten_clean() then case when pg_catalog.circle_sub_pt(
          cast(cast(null as circle) as circle),
          cast((select location from public.emp limit 1 offset 13)
             as point)) ~ cast(nullif(case when cast(null as box) &> (select boxcol from public.brintest limit 1 offset 2)
                 then (select f1 from public.circle_tbl limit 1 offset 4)
               else (select f1 from public.circle_tbl limit 1 offset 4)
          case when (select pg_catalog.max(class) from public.f_star)
                 ~~ ref_0.c then cast(null as circle) else cast(null as circle) end
            ) as circle) then ref_0.a else ref_0.a end
       else case when pg_catalog.circle_sub_pt(
          cast(cast(null as circle) as circle),
          cast((select location from public.emp limit 1 offset 13)
             as point)) ~ cast(nullif(case when cast(null as box) &> (select boxcol from public.brintest limit 1 offset 2)
                 then (select f1 from public.circle_tbl limit 1 offset 4)
               else (select f1 from public.circle_tbl limit 1 offset 4)
          case when (select pg_catalog.max(class) from public.f_star)
                 ~~ ref_0.c then cast(null as circle) else cast(null as circle) end
            ) as circle) then ref_0.a else ref_0.a end
       end as c0,
  case when (select intervalcol from public.brintest limit 1 offset 1)
         >= cast(null as "interval") then case when ((select pg_catalog.max(roomno) from
             !~~ ref_0.c)
        and (cast(null as xid) <> 100) then ref_0.b else ref_0.b end
       else case when ((select pg_catalog.max(roomno) from
             !~~ ref_0.c)
        and (cast(null as xid) <> 100) then ref_0.b else ref_0.b end
       end as c1,
  ref_0.a as c2,
  (select a from public.idxpart1 limit 1 offset 5) as c3,
  ref_0.b as c4,
      cast((select pg_catalog.sum(float4col) from public.brintest)
         as float4)) over (partition by ref_0.a,ref_0.b,ref_0.c order by ref_0.b) as c5,
  cast(nullif(ref_0.b, ref_0.a) as int4) as c6, ref_0.b as c7, ref_0.c as c8
  public.mlparted3 as ref_0
where true;

However, just like in this 40-line, 2.2kB example, the random queries generated by SQLsmith that trigger some error are most often very large and contain a lot of noise that does not contribute to the error. So far, manual inspection of the query and tedious editing was required to reduce the example to a minimal reproducer that developers can use to fix the problem.

Reduce complexity with SQLreduce

This issue is solved by SQLreduce. SQLreduce takes as input an arbitrary SQL query which is then run against a PostgreSQL server. Various simplification steps are applied, checking after each step that the simplified query still triggers the same error from PostgreSQL. The end result is a SQL query with minimal complexity.

SQLreduce is effective at reducing the queries from original error reports from SQLsmith to queries that match manually-reduced queries. For example, SQLreduce can effectively reduce the above monster query to just this:

SELECT pg_catalog.stddev(NULL) OVER () AS c5 FROM public.mlparted3 AS ref_0

Note that SQLreduce does not try to derive a query that is semantically identical to the original, or produces the same query result – the input is assumed to be faulty, and we are looking for the minimal query that produces the same error message from PostgreSQL when run against a database. If the input query happens to produce no error, the minimal query output by SQLreduce will just be SELECT.

How it works

We’ll use a simpler query to demonstrate how SQLreduce works and which steps are taken to remove noise from the input. The query is bogus and contains a bit of clutter that we want to remove:

$ psql -c 'select pg_database.reltuples / 1000 from pg_database, pg_class where 0 < pg_database.reltuples / 1000 order by 1 desc limit 10'
ERROR:  column pg_database.reltuples does not exist

Let’s pass the query to SQLreduce:

$ sqlreduce 'select pg_database.reltuples / 1000 from pg_database, pg_class where 0 < pg_database.reltuples / 1000 order by 1 desc limit 10'

SQLreduce starts by parsing the input using pglast and libpg_query which expose the original PostgreSQL parser as a library with Python bindings. The result is a parse tree that is the basis for the next steps. The parse tree looks like this:

├── targetList
│   └── /
│       ├── pg_database.reltuples
│       └── 1000
├── fromClause
│   ├── pg_database
│   └── pg_class
├── whereClause
│   └── <
│       ├── 0
│       └── /
│           ├── pg_database.reltuples
│           └── 1000
├── orderClause
│   └── 1
└── limitCount
    └── 10

Pglast also contains a query renderer that can render back the parse tree as SQL, shown as the regenerated query below. The input query is run against PostgreSQL to determine the result, in this case ERROR: column pg_database.reltuples does not exist.

Input query: select pg_database.reltuples / 1000 from pg_database, pg_class where 0 < pg_database.reltuples / 1000 order by 1 desc limit 10
Regenerated: SELECT pg_database.reltuples / 1000 FROM pg_database, pg_class WHERE 0 < ((pg_database.reltuples / 1000)) ORDER BY 1 DESC LIMIT 10
Query returns: ✔ ERROR:  column pg_database.reltuples does not exist

SQLreduce works by deriving new parse trees that are structurally simpler, generating SQL from that, and run these queries against the database. The first simplification steps work on the top level node, where SQLreduce tries to remove whole subtrees to quickly find a result. The first reduction tried is to remove LIMIT 10:

SELECT pg_database.reltuples / 1000 FROM pg_database, pg_class WHERE 0 < ((pg_database.reltuples / 1000)) ORDER BY 1 DESC ✔

The query result is still ERROR: column pg_database.reltuples does not exist, indicated by a ✔ check mark. Next, ORDER BY 1 is removed, again successfully:

SELECT pg_database.reltuples / 1000 FROM pg_database, pg_class WHERE 0 < ((pg_database.reltuples / 1000)) ✔

Now the entire target list is removed:

SELECT FROM pg_database, pg_class WHERE 0 < ((pg_database.reltuples / 1000)) ✔

This shorter query is still equivalent to the original regarding the error message returned when it is run against the database. Now the first unsuccessful reduction step is tried, removing the entire FROM clause:

SELECT WHERE 0 < ((pg_database.reltuples / 1000)) ✘ ERROR:  missing FROM-clause entry for table "pg_database"

That query is also faulty, but triggers a different error message, so the previous parse tree is kept for the next steps. Again a whole subtree is removed, now the WHERE clause:

SELECT FROM pg_database, pg_class ✘ no error

We have now reduced the input query so much that it doesn’t error out any more. The previous parse tree is still kept which now looks like this:

├── fromClause
│   ├── pg_database
│   └── pg_class
└── whereClause
    └── <
        ├── 0
        └── /
            ├── pg_database.reltuples
            └── 1000

Now SQLreduce starts digging into the tree. There are several entries in the FROM clause, so it tries to shorten the list. First, pg_database is removed, but that doesn’t work, so pg_class is removed:

SELECT FROM pg_class WHERE 0 < ((pg_database.reltuples / 1000)) ✘ ERROR:  missing FROM-clause entry for table "pg_database"
SELECT FROM pg_database WHERE 0 < ((pg_database.reltuples / 1000)) ✔

Since we have found a new minimal query, recursion restarts at top-level with another try to remove the WHERE clause. Since that doesn’t work, it tries to replace the expression with NULL, but that doesn’t work either.

SELECT FROM pg_database ✘ no error
SELECT FROM pg_database WHERE NULL ✘ no error

Now a new kind of step is tried: expression pull-up. We descend into WHERE clause, where we replace A < B first by A and then by B.

SELECT FROM pg_database WHERE 0 ✘ ERROR:  argument of WHERE must be type boolean, not type integer
SELECT FROM pg_database WHERE pg_database.reltuples / 1000 ✔
SELECT WHERE pg_database.reltuples / 1000 ✘ ERROR:  missing FROM-clause entry for table "pg_database"

The first try did not work, but the second one did. Since we simplified the query, we restart at top-level to check if the FROM clause can be removed, but it is still required.

From A / B, we can again pull up A:

SELECT FROM pg_database WHERE pg_database.reltuples ✔
SELECT WHERE pg_database.reltuples ✘ ERROR:  missing FROM-clause entry for table "pg_database"

SQLreduce has found the minimal query that still raises ERROR: column pg_database.reltuples does not exist with this parse tree:

├── fromClause
│   └── pg_database
└── whereClause
    └── pg_database.reltuples

At the end of the run, the query is printed along with some statistics:

Minimal query yielding the same error:
SELECT FROM pg_database WHERE pg_database.reltuples

Pretty-printed minimal query:
FROM pg_database
WHERE pg_database.reltuples

Seen: 15 items, 915 Bytes
Iterations: 19
Runtime: 0.107 s, 139.7 q/s

This minimal query can now be inspected to fix the bug in PostgreSQL or in the application.

About credativ

The credativ GmbH is a manufacturer-independent consulting and service company located in Moenchengladbach, Germany. With over 22+ years of development and service experience in the open source space, credativ GmbH can assist you with unparalleled and individually customizable support. We are here to help and assist you in all your open source infrastructure needs.

Since the successful merger with Instaclustr in 2021, credativ GmbH has been the European headquarters of the Instaclustr Group, which helps organizations deliver applications at scale through its managed platform for open source technologies such as Apache Cassandra®, Apache Kafka®, Apache Spark™, Redis™, OpenSearch®, PostgreSQL®, and Cadence.
Instaclustr combines a complete data infrastructure environment with hands-on technology expertise to ensure ongoing performance and optimization. By removing the infrastructure complexity, we enable companies to focus internal development and operational resources on building cutting edge customer-facing applications at lower cost. Instaclustr customers include some of the largest and most innovative Fortune 500 companies.