Comparison: DuckDB vs PostgreSQL

Modified date: Monday, June 30, 2025

Table of Contents

Introduction
General
Data Types
SQL
Storage and System
Benchmarking
Tools
Export Regulations
Internal
Internal - Optimizer
Internal - Runtime
Internal - Storage Engine

General

Feature	DuckDB	PostgreSQL	Definition
intro	DuckDB is an in-process SQL OLAP database management system. Simple, feature-rich, fast & open source.	PostgreSQL is a powerful, open source object-relational database system with over 35 years of active development that has earned it a strong reputation for reliability, feature robustness, and performance.	in their own words - but I reserved the rights to remove some bold claims like "the best", unless it is widely recognized.
vendor	DuckDB Lab	PostgreSQL Committees
initial release	2018	1998
latested version	1.3.0 "Ossivalis"	17	We don't put a release date here as the software is patching frequently. So tracking it is not much useful.
supported platforms	all major OS (Linux, macOS, Windows) and CPU (x86, ARM)	x86, PowerPC, S/390, SPARC, ARM, MIPS, RISC-V, and PA-RISC	supported OS/CPU platforms
db-engines ranking	45	4	ranks from https://db-engines.com/en/ranking (06/25)
relational?	yes	yes	Is it a relational database? (1) Most database are actually with some extensions, for example, nested data types, graph support, etc, which we usually called "multi-model". (2) Some of them are product family, meaning they have more than one database. Here we focus on the main one but explain others when needed.
open source?	yes	yes	mainly the engine code
license	MIT	PostgreSQL PostgreSQL is released under the PostgreSQL License, a liberal Open Source license, similar to the BSD or MIT licenses -- allows for commercial use and modification of the software.
price: box software	0 free download	0 free download
on-premise offering	yes	yes	if no means you can't buy "box" software from them
cloud offering	yes: via MotherDuck MotherDuck is a cloud data warehouse platform that builds on top of DuckDB, an open-source analytical database. It provides a managed, serverless environment for storing, sharing, and collaborating on data while offering features like cloud storage, a web-based SQL IDE, and hybrid query execution (combining local and cloud processing).	cloud vendors PostgreSQL’s cloud offerings are diverse, with many managed services available across major providers. For example, Amazon RDS for PostgreSQL. Technically, there are many efforts on improving PostgreSQL's native cloud implementation. The most famous one is Neon, which is acquired by DataBricks on 2025.
technical doc	https://duckdb.org/docs/stable/	www.postgresql.org/docs

SQL

Feature	DuckDB	PostgreSQL	Definition
SQL: standard complaince	high	high
Extension Mechanism	DuckDB Extensions	C programming, link with engine
max SQL length		undefined The actual limit may depends on limits of the SUT, for example, its memory/system swap settings: PostgreSQL internally use a struct named "StringInfo" to hold the SQL string, which is with MAX_INT length.	maximal SQL statement length
PL: main		SQL + PL/PgSQL	main programming lanage: most database suports SQL because SQL is a well established standard. However, each database would like to extend SQL more or less.
PL: other language support		yes	PL lanaguage other than PL/SQL, like PL/Java, PL/Rust etc
SP: max parameters		100
UDF: max parameters		100
SQL: max parameters		65535	number of parameters in a PREPARED query
SQL: query hints		GUC only	if it allows use query hints to guide the optimizer
SQL: explicit locking		yes: row, page, table level	Locking is usually an internal matter - so does it allow explicit locking? What levels do they support?
Triggers?		yes	If support triggers
Triggers: scope		tables, views, foreign tables	What objects can have triggers
Triggers: type		BEFORE, AFTER, INSTEAD OF	Types of triggers supported
Object-Relational?		yes
vector search		no native Extensions like pgvector supports it.	does it support vector search

Storage and System

Feature	DuckDB	PostgreSQL	Definition
arch: server	Embedded	C/S	Embedded or traditional C/S?
arch: run in browser?	yes	no/yes Officially not supported, but there are some good OSS projects like Pglite.	It also known as a client-side database, is a database that is stored and managed within a user's web browser, rather than on a remote server.
ACID: durability	yes	yes
Materialized View: support?	no	yes
arch: in-memory support		no
arch: Multi-master support?		no	if multi-master support?
arch: primary/read replica?		yes	if primary + mulitiple read replica supported
replication: sync/async		both	Can commits wait or w/o wait for replicas to acknowledge
replication: WAL shipping		yes	Uses write-ahead log (WAL) shipping for replication
replication: quorum-based commit		no	Multiple synchronous replicas with quorum for commit
database: max number		4,294,950,911 Note that 4G-4294950911=16385. The reason is that a database needs assign an object ID (you see the db folder is a number), which is starting from 16385 because smaller are reserved by boostrap.
tables: max number per database		1,431,650,303
partitions: methods		Range, List, Hash; limited composite (multi-level partitioning more limited than Oracle/SQL Server).	Supported partitioning strategies (range, list, hash, etc.).
partitions: global index		no	index across partitions
rows: max rows per table		32T/row_size	The actual number depends on storage etc
index: max allowed index		rows per table is restricted by the number of rows that can fit onto MAX_UINT*8K pages.	Max number of indices allowed per table
ACID		yes/yes	for DML and DDL
ACID: max isolation level		Serializable PostgreSQL implements Serializable Snapshot Isolation (SSI) in version 9.1.
ACID: max ANSI isolation level		Serializable

Benchmarking

Feature	DuckDB	PostgreSQL	Definition
any official TPC benchmarks?	no	no	The TPC benchmark includes a set of tests simulating real-world scenarios to evaluate database performance.

Tools

Feature	DuckDB	PostgreSQL	Definition
command line client	duckdb	psql	it means "sql client" for database supporting SQL. For embedded atabase, the client includes the server together.
admin(GUI)		pgAdmin

Export Regulations

Both are open source databases, so they are easy with export regulations.

Feature	DuckDB	PostgreSQL	Definition
Jurisdiction	US	US	Which country controls export
ECCN	None	None	An Export Control Classification Number (ECCN) is a five-character alphanumeric code used to categorize items on the Commerce Control List (CCL) for export control purposes. Most database may fall into 5D992.c category, "mass market encryption", which means it has some ordinary encryption related code, for example, the SSL connection code.
Eligible License Exception / CCATS	Not required	Not required There is no ECCN for open source software	A License Exception is an authorization that allows you to export or reexport items subject to the EAR without needing to obtain a specific export license, provided certain conditions are met. CCATS stands for Commodity Classification Automated Tracking System. The BIS assigns a CCATS number to products that it has classified under the Commerce Control List (CCL).
Encryption Components	SSL	SSL	Crypto functionality that triggers control

Internal

Feature	DuckDB	PostgreSQL	Definition
Implementation language	C++ The entire code base is starting from scratch, with emulation C/S layer to PostgreSQL. But implmentation is C++.	C	A DBMS may use mulitple programming languages, for example, supports its stored procedure. The major programming language used to implement the engine.
concurrency control		MVCC
MVCC: implemented?		yes	if implement MVCC for concurrency control
MVCC: rollback segment		no PostgreSQL uses VACUUM to clean up old versions: it is more difficult than in-memory database because the old versions are mixed with new versions on database pages.	if uses rollback segment (RS) to store old versions. Without RS, old versions and new versions are mixed stored, then the database engine has to find a way to efficiently drop the old versions at certain point.

Internal - Optimizer

Feature	DuckDB	PostgreSQL	Definition
CBO?		yes	if it employees a cost based optimizer
framework		System-R	System-R is more like a dynamic programming, bottom up optimizer, while Cascades/Volcano gebaseerd optimizer is more like top-down optimizer.
join order search		dynamic programming + Generic Algorithm (GEQO) It also allows to extend the join order algorithm.	How join order permutations are explored during plan generation.
stats: multi column		yes	Multi-column stats may cause storage space bloat: for example, if one dimension has 100 buckets, then three dimensions will require 1M buckets - but reducing the total number of buckets will result in reduced accuracy.
query hints?		GUC only	if it allows use query hints to guide the optimizer
plan guide?		no	Can we use plan guide to correct the plan? This is a more systematic and accurate way to repair the plan than plan hints. Especially during system upgrades, if the plan becomes worse, we can use plan guide to force a query to use a previous plan.

Internal - Runtime

Feature	DuckDB	PostgreSQL	Definition
resource management		simple: work_mem controls per-operator memory use	How execution memory is allocated and limited.
spilling support		yes	"spilling" refers to the process of writing temporary data or intermediate results of a query to disk when the available memory is exhausted. This is crucial for handling large datasets or complex queries that require more memory than available.
model		Volcano + push for parallel run	The operator scheduling model: pull-gebaseerd (Volcano), push etc.
support intra-parallel query?		yes	meaning a single query can utilize multi hardware threads to run it
adaptive execution (AQP)?		no	Traditionally, after the optimizer determines the plan, the runtime must execute it completely without any room for adjustment, such as which of the two tables should be built. The adaptive method allows the runtime to make some adjustments based on the actual situation, and the optimizer must also prepare for this uncertainty, such as preparing an alternative plan.
vectorization		no native Several extensions allowed it. For example, pg_duckdb and Hydra.	Speed up OLAP queries using vectorized execution. A vectorized runtime exchange a bunch of rows between iterators, and these rows are physically sotre with column-oriented order.
iterator: join methods		all 3	Hash Join (HJ), Sort-Merge Join (MJ) and Nested loop Join (NLJ) are 3 major ones
Error: out-of-range and overflow		abort the transaction	To maintain atomic requirement of ACID, database engine usually fail the statement and abort the transaction.

Internal - Storage Engine

Feature	DuckDB	PostgreSQL	Definition
ACID: durability mechanism	WAL	WAL	How is it implements durability. Database's classic way is write ahead logging (WAL).
Page Size		8K default configurable with the --with-blocksize=BLOCKSIZE option	The physical disk block size for storage and I/O.
Storage Compression		TOAST compression (within a row)
Row Overflow / Large Row Handling		Row within a page and large columns pointing to TOAST storage	How rows larger than page size are handled. Usually a row must be within one page, but big columns can only save a pointer with the row and pointing to the actual storage. This is also why there is a limit of maximal columns per table.
Index: primary cluster		Heap storage by default; CLUSTER command for static clustering - if modification happens after, order is not maintained.	Whether physical rows are ordered by primary index.
Index: secondary		B+Tree, GIN, GiST, SP-GiST, BRIN, Hash, Fulltext (via extensions).	Supported index types beyond the primary key.
index: column store as primary		No native support; available via extensions (e.g. cstore_fdw, citus).	if column-store as primary storage supported
index: column store as secondary		No native support; available via extensions (e.g. cstore_fdw, citus).	if column-store as secondary index supported
partition table		declarative partition table is implemented via inheritance table, since V11	Partitioning capabilities for large tables.
partition table: prunning		mostly at compile time, some runtime	Ability to eliminate partitions at query compile time and runtime