We tried using Postgres with large analytics at my previous company https://threekit.com but it is an absolute pain. Basically we started to collected detailed analytics and thus had a rapidly growing table of around 2B records of user events during their sessions. As it grew past a 500 million records it turned out to be impossible to query this table in any thing close to real-time - it was basically untouchable because it was so slow.
I know I could have used some type of daily aggregation combined with a weekly aggregation, etc to roll up the data incrementally. A dev tried this and yeah, it hide the slow queries but then it became inflexible in terms of reporting. And writing and maintaining these cronjobs is a lot of work.
But then I tried using BigQuery on my https://web3dsurvey.com side project and I just recorded raw data and then wrote queries to do real-time aggregation (with a caching layer) in a large variety of ways. And it was near instant and easy and also very cheap.
So then I deployed this strategy over at Threekit.com and it also was easy and fast and cheap. Even more flexible queries than Web3DSurvey at basically no developer cost or infrastructure costs. No more need for aggregating cron-jobs or trying to decide ahead of time how users want to slice and dice the data. Real time, flexible queries on the fly via BigQuery is the way.
Also BigQuery bill for https://web3dsurvey.com is like $0.25 month and it is dealing with millions of records in its 3 month window of stored data. Where as just running the cheapest Postgres SQL server on Google Cloud is like >$25/month and it is a slow one.
I would never go back to traditional SQL for analytics - it was hard, slow, expensive and inflexible. Worst of all worlds.
Nobody in their right mind would argue that Postgres without columnar storage is good for analytics. However, it looks like these extensions can be quite decent, at least at queries (you might still run into write performance due to Postgres' OLTP architecture, but I've never benchmarked it). In OLAP terms the size of your data is tiny and this would probably work just fine.
Personally I'm a huge fan of dedicated databases like ClickHouse. BigQuery gets very expensive as your data and query volume grows.
Plain Postgres tables hosted on Google Cloud SQL on a decent machine.
> Nobody in their right mind would argue that Postgres without columnar storage is good for analytics.
Which is what I am saying. Standard means without extensions.
> Personally I'm a huge fan of dedicated databases like ClickHouse.
Where do you run it? What type of machine? Can it be server less?
> BigQuery gets very expensive as your data and query volume grows.
I didn't experience this, it has been hard for me to make BigQuery cost a lot. Storage costs were basically nothing and we just set the clustering/partition on DB creation to be users/sessions/events and dates respectively. It aligned really well with general query patterns.
At my company we self-host ClickHouse, but if we started out today we'd definitely adopt ClickHouse Cloud. One big advantage of CH is that tables can be backed by ordinary cloud storage like S3 and GCS, which is incredibly cheap. BQ has support for external Iceberg tables in GCS, but (from what I recall) no write support.
We are shuffling several TBs/week into BigQuery, and that is extremely expensive. Querying this amount of data (billions of rows per day) is also very expensive unless you purchase flat pricing.
If your needs are really modest, then BQ is definitely very reasonable.
While these new Postgres extensions will make Postgres more competitive, CH is built from the ground up for these kinds of analytics workloads in a way that Postgres isn't, e.g. in terms of scalable data ingestion.
CH also has a rich set of features like materialized views, aggregate function types, and dictionaries, none of which have counterparts in Postgres of these extensions.
AFAIK, the cheapest Postgres server on GCP is very expensive compared to the usual Postgres installation (price/performance).
[0]: https://github.com/ghtorrent/ghtorrent.org/blob/master/gclou...
For that, I found Digital Ocean to be very reasonable.
This is a solved problem, and it seems the technical folks over there lacked the skills to make it work. Having indexes is just the tip of the iceberg. Composite indexes, partitioning, sharding, caching, etc, can lower reads to a few seconds on disk.
Or just use BigQuery and it is works, it is cheaper to run (by 10x to 100x) and can be done by a junior dev rather than a PhD in Database configuration.
I prefer simple solutions though - I also hate Kubernetes: https://benhouston3d.com/blog/why-i-left-kubernetes-for-goog...
For this workload, having a columnstore version of your table will help.
DM us: https://join.slack.com/t/mooncakelabs/shared_invite/zt-2sepj.... We can help.
But if BigTable just solves the problem it seems the way to go
PostGres is popular for a reason, it’s ACID unlike BigTable
You may run into these problems later on, you may not
What do you mean by "standard" SQL? Were you using an OLTP or OLAP db when you faced these difficulties? Also, what makes BigQuery not a "standard SQL database" -- they use GoogleSQL, true, but many DBs have slight SQL variants.
No extensions, just the default Postgres Cloud SQL offering from Google Cloud.
> Also, what makes BigQuery not a "standard SQL database" -- they use GoogleSQL, true, but many DBs have slight SQL variants.
I view BigQuery as not a standard DB because it really doesn't do transactions, rollbacks, foreign key constraints, indices, stored procedures, etc. I view it as a non-SQL database pretending to be SQL for query convenience.
Parts of SQL DDL is implementation specific. The SQL standard hardcodes the assumption of (by modern standards) a small-scale system and therefore will deviate significantly on large-scale systems. There is no practical way to make this "standard", you have to understand the implementation it reflects.
The core idea of mooncake is to built upon open columnar format + substitutable vectorized engine, while natively integrate with Postgres.
So it is indeed closer to BigQuery (especially the newer bigquery with iceberg tables) than a 'standard SQL database'. It has all the nice properties of BigQuery (ObjectStore-native, ColumnStore, Vectorized execution...) and scaling is also not impossible.
I've worked with a 5B row table on Snowflake with (maybe) no indexes, and while somewhat slow you could still run reasonable queries on it in a minute or two.
I mean, I don't know what you call "close to real time", and what kind of query you did, but I have Postgres serving requests from a 20B rows table just fine, with some light tweaking of indexes and partitions (I'm by no means a DBA).
Unless we are digging much deeper than that in the specifics of the data and queries, I don't think any conclusion can be taken from either experience.
You can do fast-twitch analytical queries (sub-second, a few seconds at the tail) on complex operational data models where you are inserting millions of records per second on 100B record tables while those queries are running. Just not on Postgres.
The biggest table contains 30B records. A query that uses a B-tree index completes in a few microseconds.
EXPLAIN ANALYZE SELECT * FROM table_name WHERE id = [ID_VALUE];
Index Scan using table_name_pkey on table_name (cost=0.71..2.93 rows=1 width=32) (actual time=0.042..0.042 rows=0 loops=1)
Index Cond: (id = '[ID_VALUE]'::bigint)
Planning Time: 0.056 ms
Execution Time: 0.052 msBut that's the way to go, isn't it?
Once you've optimized for the more obvious filters like timestamp and primary key, like using partitions to avoid using indexes in the first place, you're left with the situation where you need to aggregate over many gigabytes of data and an index doesn't help since your query is probably going to touch every page within the filtered partitions.
You can solve some of these problems in Postgres, like partitioning, but now you're stuck with random I/O within each page to perform non-SIMD aggregations in a loop. This approach has a ceiling that other implementations like ClickHouse do not.
Analytics generally means a full table scan or similar for each query, because you're e.g. taking an average of every value.
You're talking about indexes which are to optimize retrieving a handful (or few thousand) of rows.
Analytics isn't typically something that needs real-time capabilities, for one.
> a rapidly growing table [emphasis mine]
I think I see part of the problem here. If you had a single table, that means it's completely denormalized, so your schema probably looked something like this (or wider):
CREATE TABLE UserEvent (
id UUID PRIMARY KEY,
user_id UUID NOT NULL,
user_ip_address TEXT NOT NULL,
user_agent TEXT NOT NULL,
event_data JSONB,
created_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
event_type TEXT
);
CREATE INDEX UserEvent_user_id_idx ON UserEvent (user_id);
CREATE INDEX UserEvent_created_at_idx ON UserEvent (created_at);
A UUID, if stored as its native type (or BINARY(16) in MySQL - don't sleep on this, MySQL folks; it makes a huge difference at scale) is 16 bytes. That's double the size of a BIGINT, and quadruple the size of an INT4. Also, unless you're using UUIDv7 (or UUIDv1, but no one does), it's not k-sortable. Since Postgres doesn't cluster tuples around the PK [yes, I know all indices in Postgres are technically secondary] like MySQL/InnoDB does, this doesn't immediately thrash the B+tree in the same way, but it does thrash the visibility map, and it does bloat the WAL. There are various arguments for why you shouldn't use a monotonic integer as a surrogate key, but IMO they're largely overblown, and there are workarounds to not publicly disclose it.
IPv4 addresses, stored in dotted-quad as a string, are a maximum of 15 characters, storing in 16 bytes as TEXT or VARCHAR. If stored instead in the Postgres native INET type, that drops to 7 bytes, plus you get built-in validation. If you had INT4 UNSIGNED available (as MySQL does natively), you could even store them in their numeric form and save another 3 bytes, though you lose the validation.
User Agent strings are huge, usually over 100 bytes. They're also not that unique, relatively speaking. Even if you need to know the patch version of the browser, anyone with a browser doing automatic updates is going to stay more or less in sync. The point is this could easily be a lookup table, with either a SMALLINT (2^15 - 1 maximum values, or 2^16 - 1 if you use unsigned values; possible with an extension in Postgres) or an INT (2^31 -1 maximum values) as the PK.
Not going to touch on JSON objects because the things you might want to know are endless. TOAST and de-TOAST can be slow; if you need low latency, you should normalize your data.
There may or may not be extracted scalars, which can be beneficial during queries. Again, though, lookup tables (or even native ENUM types, if the values are limited) are crucial at scale.
As it stands, the table will have an average row size of 664 bytes (assuming an average of 12 bytes stored for the IP, 100 bytes stored for the UA, 500 bytes stored for the JSONB, and 12 bytes stored for the event type). That's 332 GB for 500,000,000 rows. You could shave a couple of bytes off by aligning columns [0], which saves 1 GB. If the IP addresses and UA strings were lookup tables, each with an INT4, that saves 104 bytes per row. If you made the PK for the table a BIGINT, that saves another 8 bytes per row. The total savings between column alignment and basic normalization is 114 bytes per row, or 57 GB.
This doesn't touch on the indices, either. If you're using PG 13+, you get B+tree de-duplication [1] for free, which can help with some denormalized data, but not if you have anything with high cardinality, like a timestamp, or a UUID. With lookup tables, you would of course need to index those FKs (whether or not you're enforcing constraints), which adds some size, but is still a huge net gain.
> I know I could have used some type of daily aggregation combined with a weekly aggregation, etc to roll up the data incrementally. A dev tried this and yeah, it hide the slow queries but then it became inflexible in terms of reporting. And writing and maintaining these cronjobs is a lot of work.
And shifting your entire analytics workload isn't a lot of work? Between ROLLUP [2] and MATERIALIZED VIEW [3], which can automatically refresh itself with a cron, this doesn't seem that burdensome.
> Also BigQuery bill for https://web3dsurvey.com is like $0.25 month and it is dealing with millions of records in its 3 month window of stored data.
Then you're in the free tier (<= 1 TiB/month of processed data), because after that it's $6.25/TiB. Also worth noting there is a massive difference between millions of rows and billions of rows. The former can be handled by practically any RDBMS on any hardware, with a completely unoptimized schema. The latter requires some thought if you want it to be performant.
This isn't at all to say that specialized DBs don't have their place, because they absolutely do. If you need a KV store, use a KV store, not an RDBMS. If you need OLAP, use something designed for OLAP. The difference is scale. At startup or side project scale, you can easily do everything (including pub/sub) with an RDBMS, and if you put thought into its design and usage, you can take it a lot farther than you'd think. Eventually, you may hit a point where it's counter-productive to do so, and then you should look into breaking tasks out.
The issue I see happening time and time again is devs have little to no expertise in DBs of any kind, but since everyone says "Postgres is all you need," they decide to use it for everything, except they don't know what they're doing. If you do that, yeah, you're gonna have problems fairly early on, and then you'll either throw your hands up and decide you really need a bevy of specialized DBs, caches, and message queues (which introduces a lot of complexity), or you'll vertically scale the DB. If you choose the latter, by the time you hit scaling limits, you're easily spending $25K/month on the DB alone. If you opt to hire someone with DB expertise at this point, you'll spend about that if not more in personnel costs, and not only will it take them weeks if not months to unravel everything, your devs will be constantly complaining that queries are now "too complex" because they have to do some JOINs, and they're being told to stop chucking everything into JSON. If instead, you took at most a week to learn some RDBMS basics by a. reading its manual front-to-back b. hands-on experience, trying things out you could almost certainly get much farther on much less.
[0]: https://www.enterprisedb.com/blog/rocks-and-sand
[1]: https://www.postgresql.org/docs/current/btree.html#BTREE-DED...
[2]: https://www.postgresql.org/docs/current/queries-table-expres...
[3]: https://www.postgresql.org/docs/current/rules-materializedvi...
For analytics on transactional data, it looks like you'd still need to use logical replication (https://github.com/Mooncake-Labs/pg_mooncake/issues/90). Logical replication is somewhat similar to an ETL/CDC experience, though it's more Postgres-native. Managing logical replication at a production grade isn't trivial — it's quite common for customers to use PeerDB for homogeneous replication due to performance, manageability and observability issues with logical replication.
One potential value of extensions is Postgres compatibility for your queries. However, I'm curious about the SQL coverage on columnar tables — for example, advanced constructs like partitioned tables, prepared statements, correlated subqueries, RCTEs, triggers, and more. While it seems there’s ongoing work to improve compatibility, achieving full Postgres compatibility is challenging. In some cases, you might need to fall back to Postgres (instead of DuckDB), sacrificing performance for certain queries.
The biggest challenge we faced at Citus was the constant trade-off between Postgres compatibility and performance — both are complex and constantly moving targets. This was the key reason why Citus couldn't compete with purpose-built databases like Snowflake, ClickHouse, SingleStore. While DuckDB didn’t exist when we built Citus, it's still fundamentally different from Postgres. Even though the DuckDB dialect is somewhat similar to Postgres (as is Snowflake’s), retrofitting two distinct databases — each evolving independently for world's best performance — isn't trivial.
In short, relying on ETL (logical replication) without providing full Postgres compatibility raises the question: is it better to lean on a purpose-built database with a laser focus on performance, or adopt a more generalized approach?
Anyway, I come from a different school of thought — using the right tool for the right job. That said, I love seeing all the progress and evolution in the Postgres community — Postgres will always be my first love!
Love your work on PeerDB and it's inspiring the evolvement of pg_mooncake (logical replication will be the killing feature for V2)
The core idea of mooncake is to built upon open columnar format + substitutable vectorized engine, while natively integrate with Postgres:
1. For small devs, we allow the whole stack to be embedded as a Postgres extension for ease of use
2. For enterprise, our stack is also purpose-built stack similar to PeerDB + ClickHouse, not a more generalized approach
We allow a gradual transition from 1 to 2.
1, makes sense.
On 2, I understand your thinking around purpose-built — but you're retrofitting an analytical database into a transactional database without fully supporting all the features (both in terms of functionality and performance) of either. It's really hard to be truly purpose-built this way. As a result, users might not get the best of both worlds.
PeerDB is different. We keep Postgres and ClickHouse separate and just move data reliably between them. Users get to query Postgres and ClickHouse in isolation and make the best of each of them.
Anyway, keep up the good work! Just wanted to share some challenges we've seen before when building an analytics extension (Citus), particularly around chasing both Postgres compatibility and performance.
Logical replication with mooncake will try to create a columnar version of a postgres heap table, that can be readable within postgres (using pg_mooncake); or outside postgres (similar to peerdb + clickhouse) with other engines like duckdb, StarRocks,Trino and possibly ClickHouse.
But since we can purposely build the columnstore storage engine to have postgres CDC in mind, we can replicate real-time updates/deletes(especially in cases traditional OLAP system won't keep up).
The focus of mooncake is to be a columnar storage engine, that natively integrate with pg, allowing writing from pg, replicating from pg, and reading by pg using pg_mooncake. We want people to use other engine to read from mooncake, and here they are effectively stateless engine, that's much easier to manage and avoids all data ETL problems.
Spoiler-alert: Mooncake will be supporting HTAP use-cases soon, and it comes with better trade-offs: keep your OLTP postgres tables as is, and mooncake adds analytics capacity to those table on up-to-date data.
So is it Postgres or DuckDB that cracked the analytics top ?
Fast analytic databases need two key things: columnar storage and a vectorized execution engine. We introduce a columnstore table access method in Postgres with data stored in Parquet) and execute queries on those tables using DuckDB.
By leveraging DuckDB's execution engine, we avoid reinventing vectorized query execution while keeping everything managed through Postgres.
More on our architecture: https://www.mooncake.dev/blog/how-we-built-pgmooncake
We blogged about this: https://www.mooncake.dev/blog/duckdb-parquet
That doesn’t seem like something that makes sense to in DuckDB.
> pg_mooncake leverages Postgres to store table metadata, eliminating the need to query an external catalog (as DuckDB would require when scanning external Iceberg or Delta Lake tables). Instead, it directly queries the Postgres table to identify which Parquet files to scan, reducing overhead.
> Additionally, pg_mooncake stores detailed Parquet metadata for each file in Postgres, including column statistics for every row group. These statistics, such as MIN/MAX values for each column, are aggregated to optimize scans. When querying a columnstore table, pg_mooncake uses these statistics to filter out data files and row groups that cannot produce matches.
It might not even be possible to install x on some hosted pg servers.
It is great that you can expand pg and that you and others have but I don't rate it as high as a buil-in solution.
> Fast analytic databases need two key things: columnar storage and a vectorized execution engine. We introduce a columnstore table access method in Postgres with data stored in Parquet) and execute queries on those tables using DuckDB.
That was always the culprit for me compared to AWS/Athena and BigQuery. They are dirt cheap on analytics workloads when you can parallelize the calculations to 100 CPUs without really paying any extra.
With postgres you are stuck with linear cost for scaling up number of CPUs, so everything is slow anyway.
Good point! Normally for postgres extension it won't be solvable, but for mooncake it is actually not the case!
The core idea of mooncake is to built upon open columnar format + substitutable vectorized engine, while natively integrate with Postgres.
So right now it is using duckdb within postgres to run the query, but we can and we will support ad-hoc using other 'stateless engines' like Athena, StarRocks or even spark to run a big query.
I'm trying to understand if there is any potential performance impact on the OLTP workload by including the OLAP in the same process.
In order to have a viable business model they need to create value for users. Users are intelligent and will not even consider trying something that has no exit path, should the company disappear.
What every company hopes to have is customers who are thrilled with the value they deliver and tell everyone what a great deal it is so the company can grow.
What no company hopes for is to end up like hashicorp, where they end up spending more investment $ than the value they provide, never achieve profitability, and eventually just piss everyone off, and everyone is trying to make the best of a leveraged situation, they end up having to pull the rug. The user's leverage in that situation is something like opentofu, made possible by the license, same as what's being offered here.
The price of the big columnar dbs is very high, so there's a lot of ground to capture / value to arbitrage/offer in this space - as evidenced by other comments in this thread, how fast the benchmarks are changing, the likelihood of memory prices coming down, etc.
Aside from that, you have to wonder big picture if the AI space will put significant downward pressure on memory and compute prices with everyone wanting to run local LLMs, might change some fundamental tradeoffs made in db systems. If in 10 years I can get a phone with a 1 TB of fast access RAM to ask siri how to sauté mushrooms, what will 99% of us need columnar store for?
Mooncake is built upon open-table formats and substitutable query engines. So it don't need to be just a postgres extension.
PG_mooncake will stay open-source under MIT, for small devs where everything fits in their postgres, we hope everyone to enjoy the !
And we would love to help companies outgrown postgres to transition into the modern stack with postgres + mooncake potentially outside pg + iceberg/delta.
Unfortunately using an extension that's not "part" of postgres (like pg_stat_statements) is not trivial for most people since there are both technical and legal issues.
It should be fairly trivial to get started
Others, like Citus, are licensed under the AGPL, which is basically an even stronger version of the GPL that requires anyone hosting a service to provide their modifications to the source code. While this doesn't strictly rule out use in e.g. RDS it does basically say that Amazon would be required to open source any changes they make to get it working in RDS, which is a burden they don't seem interested in taking on, and which might reveal details about how their services work under the hood which they are generally fairly cagey about doing.
The AGPL is also viewed even more suspiciously than the GPL by corporate lawyers who fear it may lead to virally forcing the company to reveal all of its source code. This fear is probably unfounded, but all copyleft licenses are primarily about expanding end-user freedom and so are treated as undesirable in most businesses that make money from selling software.
Even if the extension has a liberal licence eg Apache or MIT nothing guarantees that the licence won't change in the future.
There are some still low-hanging fruit optimizations we're working on (https://github.com/Mooncake-Labs/pg_mooncake/issues/82), and we'd love to get back to the Top 10.
We embed DuckDB in our extension as the vectorized execution engine on columnstore tables.
Why rebuild a vectorized execution, when DuckDB has a lot of GREAT properties for this workload (similar syntax to Postgres, embedability).
Here's our architecture: https://www.mooncake.dev/blog/how-we-built-pgmooncake
We implement segment-elimination to do so. We even blogged about it: https://www.mooncake.dev/blog/duckdb-parquet
1. Our extension is fully open-source. I believe they've open-sourced bits of their stack. 2. We are unopionated about open table formats: Iceberg and Delta. 3. In v0.2, we will support small write workloads. This will open up time-series and HTAP workloads.
It's a generally available (very solid) product powering some large production workloads, with fully transactional Iceberg, and auto-compaction. All SQL queries and almost all Postgres features are fully supported on Iceberg tables.
We are also seeing interesting patterns emerging with the ability to load/query csv/json/parquet/shapefile/... directly from S3 in combination with pg_parquet and pg_incremental. For instance, incrementally & transactionally loading CSV files that show up in S3 into Iceberg, or periodically exporting from Postgres to Parquet and then querying with data warehouse.
In November last year we added Iceberg support and managed storage, so it became a full data warehouse experience with writes and transactions and we renamed it to Crunchy Data Warehouse.
Keep your OLTP as it. Deploy Mooncake with logical replication into columnstore tables, and get DuckDB like perf.
A big part of this is not replacing your OLTP.
However, the Venn diagram does not tell me anything useful, so I agree with the sentiment. Would love it if someone could explain what mooncake is good for and what its strengths and weaknesses are vs Timescale.
pg_mooncake (&crunchyData) is implementing columnstore tables in postgres, so you can actually use postgres as a data-warehouse (to ingest/ update and run analytics queries)
This will be fixed in v0.2. Stay tuned, we will have some announcements here coming soon.
With what we're working in v0.2, we re-architecture some of our dependencies there, and will be able to scale out and support Citus.
1. pg_duckdb is great for querying existing columnar files (parquet) in Postgres.
2. Our focus is on helping you write and query your existing Postgres table into a columnar format.
We spent most of our time on building the table access method for columnstore tables.