How To Load One BILLION Rows into an SQL Database
📝 Get my free SQL Cheat Sheets: www.databasest...
🎓 Learn and become confident in SQL: databasestar.m...
Have you ever wondered how to import one billion rows into an SQL database? Or how long it would take?
In this video, you'll see how to generate a file with 1 billion rows, how to import it into four different SQL databases, and how long it takes in each database.
⏱ TIMESTAMPS:
00:00 - The challenge
00:31 - Setting up the data
02:08 - MySQL
04:08 - Postgres
05:31 - SQL Server
07:35 - Oracle method 1
09:33 - Oracle method 2
🔗 VIDEO LINKS:
Gunnar Morling's blog post: www.morling.de...
1BRC GitHub repo: github.com/gun...
SQL scripts used in this video: github.com/bbr...
Commands to clone the repo and generate the file:
git clone github.com/gun...
java -version
mvnw verify
sh create_measurements.sh 1000000000
Specs of the computer used in this video:
Lenovo Ideapad 5 laptop
CPU: Intel i5 2.4 GHz
RAM: 16 GB
OS: Windows 11
Пікірлер: 108
Thanks for all of the comments everyone! I'm glad a couple of issues have been pointed out. I'll address them when I redo the process in a follow-up to this video, including: - use SQL Server Developer Edition rather than Express to avoid the database size limit and import all 1B rows - use another version of Oracle to also avoid the database size limit and import all 1B rows - consider optimising the queries or changing some parameters to further improve the performance of the SELECT
If doing this with 250 million rows and multiplying by 4 was a correct approach, then for all databases you could just use 1000 rows and multiply by a million.
@jpsolares
Ай бұрын
Perhaps consider using the Developer Edition, as it is equivalent to the Enterprise Edition. Since the project is not intended for production purposes, there should be no issue in utilizing it.
@ahmettek315
Ай бұрын
@@jpsolares We all know it. What I am saying is, you can't test something doing with 250 million rows and multiplying by 4 and then claim it is equal to doing with 1 billion rows. Looks like you understood as if I were asking how to do it.
@DatabaseStar
Ай бұрын
Good point! The only reason I used 250m for Oracle and SQL Server was because of the size limitations mentioned in the video. In hindsight (and from the other comments), I could have used the SQL Server Developer edition to load the full 1 billion rows. I could also have used a different edition for Oracle just for this experiment.
@ahmettek315
Ай бұрын
@@DatabaseStar No problem. I later read about it a bit, and it was indeed a challenge where you could use any kind of optimization you can. In my view, that made your SQL server and postgreSQL samples void for me. I don't use mySQL and Oracle, but at least I know your sample is void for mySQL as well. You are doing the uploads and querying just how a beginner would do as well. No tricks, no optimizations, no use of extensions, CLR etc. I could create better versions for postgreSQL and SQL server but only when I have spare time and\or I think it is worth doing it.
@DatabaseStar
Ай бұрын
Yes, that's true, I mention in the video I don't make any optimisations such as indexes or adjustments to the process. The default settings are used on purpose. There are definitely ways to make it faster in each database!
Time needed for a certain query doesn't grow linearly in respect to the size of data 6:00
@DatabaseStar
Ай бұрын
Yes I completely agree. I mentioned in the video it's not a true comparison as I was unable to load the full 1B rows into Oracle and SQL Server.
in SQL server, if you change the auto grow to 10,000 from 64 it will run much faster for both data and log files. Also SQL server has external tables as well.
@lucca5101
Ай бұрын
As does Postgres, but he only used that feature with Oracle. Every other DB he copied the data. Totally not pushing Oracle…
@DatabaseStar
Ай бұрын
Thanks for the tip, I wasn't aware of that! I left all of the databases at their default settings but this would have been a good one to change to avoid the issue.
@DatabaseStar
Ай бұрын
No, I'm not pushing Oracle, and I left all of the settings as the defaults. I did an additional method for Oracle because I was aware of its ability to use an external table.
INSERT or BULK-INSERT do not have a linear data insertion time, but an exponential one. Therefore the test should be repeated using 250,000 records for each database (in the comparison of no-cost databases). Oracle and MSSql in the paid versions have very respectable performances as they no longer have RAM or CPU core limits
@DatabaseStar
21 күн бұрын
Thanks for the tip, yeah that's right. Others have pointed that out as well. In hindsight I should have used different versions of Oracle and SQL Server to avoid the size limits, so I could import the billion rows in all vendors.
To solve the growth problem in SQL Server, when creating the database, you must click on the ... and choose the growth in percentages, this prevents you from having to indicate a value that may be short (that is, not establishing a growth limit) and thus you solve the problem of the size of the database
@DatabaseStar
Ай бұрын
Good tip, thanks for sharing!
Please include sqlite
@DatabaseStar
Ай бұрын
Good idea
Loading 250m rows 4 times is not equal to loading 1billion rows. The experiment is really flawed. You need to use 250m rows for all. looking forward to another test with same yardstick. Kudos!
@DatabaseStar
Ай бұрын
Thanks! Yeah a few people had pointed that out, and I mentioned it a couple of times in the video. I plan on doing a follow-up video for it.
You can freely and legally use Developer edition for an MS Sql Server. As long as there no production data loaded, the Dev edition is free.
@DatabaseStar
Ай бұрын
Thanks for the tip! I should have used that edition, in hindsight, for a better comparison.
In a real world scenario, likely you wouldn't just keep that data as two columns city, temperature. Checking the original challenge, it is more about parsing data. Using plain streaming with languages (C series, Java, Go, Rust, ...) is expected to beat any database that way IMHO. They don't need to store all the values, just min, max and average (aka Sum and count).
@DatabaseStar
Ай бұрын
Yeah the results of the Java challenges were impressive. One of the points of this video is that databases are great at processing data, so I wanted to see how the different databases handled it.
@ahmettek315
Ай бұрын
@@DatabaseStar Indirectly I started to dig what was that about, those results were not from a file on disk but from a RAM disk and the machine used is an AMD EPYC 7502P with 32 cores and 128 Gb memory! That CPU is supposedly more than 5 times better than yours and is a server CPU. Directly working from memory on a beast hardware those timings is not fair to compare with yours. Maybe should also run one or more of their code to see what they score on the same hardware. (Trying to allocate time for a postgreSQL and\or Go check)
It is interesting to see a real test for the deference in loading/select speed between DBs, and how out of the box Oracle numbers (select) are more than13x times faster than postgres and mysql I wander if there are way to tweaked some parameters in Postgres/Mysql (especially for Mysql) to decrease those number and make them comparable to Oracle. Thanks a lot for the demonstration, it is a grate video
@DatabaseStar
Ай бұрын
Glad you liked the video! Yes I thought the test was interesting as I was doing it. I think there would be ways to tweak some parameters on the server to improve the import process. The default settings may not be the fastest. Also there are indexes that could have been added after the import which could have also improved the query performance.
@ahmettek315
Ай бұрын
It is only so, the way he did. I don't use mySQL but look at vitess project for example. AFAIK it is what made possible youtube to store their videos on mySQL. You may also check with PlanetScale.
Oh I love Oracle! I've been using for years and fascinated with all its features, functions and properties. Its direct path loading feature is also awesome.
@DatabaseStar
24 күн бұрын
Yeah, it does have a lot of features.
bcp, or load a mdb on a non-production server then copy the file and sp_attachdb it. you’ll need to stop the service, then attach it. I can think of a half dozen ways to do this, depending on business constraints.
@DatabaseStar
24 күн бұрын
Good to know there are many other ways to do this
Have you used Clickhouse or DuckDB??
@DatabaseStar
Ай бұрын
No, I haven't used them actually.
How about SQL*Loader in Oracle?
@DatabaseStar
24 күн бұрын
Good idea!
Interesting, thank you for sharing. For the postgres part, do you think the pagination negatively affected the results?
@DatabaseStar
24 күн бұрын
No problem! I think it did, actually, based on a couple of other comments. In hindsight I should have shown all results at once in the query to get a more accurate number.
You should extend your test for databases like: SAP HANA, IBM DB2 , Snowflake, Google BigQuery, MariaDB, SQLite
@DatabaseStar
17 күн бұрын
Good idea! I don't have any experience with those databases unfortunately.
@krzysztofcierpia4853
14 күн бұрын
@@DatabaseStar You can find free/community edition for SAP Hana (Express) and DB2 (community edition )
Much Helping
@DatabaseStar
Ай бұрын
Thanks!
concurrency is a critical factor (one of the many others) in the context of this scenario. based on my observations, SMP databases exhibit subpar performance when managing such workloads in a production setting. in contrast, MPP databases such as teradata, snowflake, and netezza offer a more suitable solution. a decade ago, I was involved in an oracle to teradata migration project, where we conducted parallel testing by executing identical queries on both database platforms. the disparities we encountered were stark and significant.
@ramsundararaman6615
Ай бұрын
Tell us more about it, please!
@DatabaseStar
Ай бұрын
Good point, yes I imagine concurrency could impact the performance here.
You can get the Developer Edition of SQL Server, which is free
@DatabaseStar
Ай бұрын
Thanks for sharing! In hindsight, I should have done this instead of using Express, because it would have been a fairer comparison.
What about comparing this to SQL serverless endpoint using an external table. This can be done in Synapse or Fabric. That would be fair compare to Oracle External table. Otherwise insightful. Thank you.
@DatabaseStar
10 күн бұрын
Good idea, I don't have any experience with those so I didn't include them
Great video. But it got me curious what about nosql databases? I know they are faster than sql but how much? Can you make a video for nosql databases? Like mongo, cassandra etc.
@flybyray
Ай бұрын
Or just use unix tools efficiently directly and/or withinOrAsPipe to the generator script itself 😂
@DatabaseStar
Ай бұрын
Thanks! They could be faster, but I don't have a lot of experience with NoSQL so I just focused on SQL for this video.
Just for curiosity, can you share your machine's specs? From create...sh timing it looks like you have a faster machine than mine but would be good to know.
@DatabaseStar
Ай бұрын
Sure! Here are the specs: Lenovo Ideapad 5 laptop CPU: Intel i5 2.4 GHz RAM: 16 GB OS: Windows 11
for sqlserver you can use developer edition
@RadityoPrasetiantoWibowo
Ай бұрын
you should use 1. sqlserver developer edition 2. you should monitor memory usage and CPU usage also 3. you should use same client tools like dbbeaver, or minimum client like CLI version of each vendor good work!
@DatabaseStar
Ай бұрын
Thanks! Yeah a few other commenters have pointed that out, and in hindsight I should have done that.
Your Postgresql select benchmark was also not correct. It wasn't 27 minutes it was much lower . So the issue was the query already returned all the results. It was your dbeaver which was paginating the result
@DatabaseStar
24 күн бұрын
Oh thanks for pointing that out, that's good to know. I'll look at it again as it seemed like DBeaver took several minutes each time it loaded a new page.
GREAT CONTENT🔥
@DatabaseStar
Ай бұрын
Thanks!
One of the best channels on database
@DatabaseStar
Ай бұрын
Thanks!
… some of the Oracle comparison-min-max-times are only for fetching 200 rows!!! … and why not loading 500m rows and multipl. x2 ?
@DatabaseStar
Ай бұрын
Good point! I tried to load 500m rows but that also exceeded the maximum size of the database (10 GB).
So, what's your hardware configuration, just a laptop?
@DatabaseStar
Ай бұрын
Yeah, it's my Windows laptop: Lenovo Ideapad 5 laptop CPU: Intel i5 2.4 GHz RAM: 16 GB OS: Windows 11
Clearly trying to push Oracle. Why not use Postgres to directly read the text file for a better comparison?
@DatabaseStar
Ай бұрын
Good idea! In my research I couldn't find a way to read a text file directly using Postgres. Do you know how it could be done? I'm not trying to push Oracle at all. My preferred database to use is actually Postgres :)
what if you index the city names, how fast would the queries be?
@DatabaseStar
Ай бұрын
They would probably be quite a bit faster!
Why you didn’t use sql server developer edition??? to avoid the express problems
@DatabaseStar
Ай бұрын
Good point! In hindsight I should have done this, and it would have been a better comparison.
How fast would it be to load all the data into an in memory Database and query it?
@Nachtaktiv2000
Ай бұрын
in memory will only be a bit faster when loading, but up to 1000x faster when querying
@DatabaseStar
Ай бұрын
Good question. I agree with what Nachtaktiv mentioned, the benefits would be when querying.
You should use SQL Server for developers edition, it's not limited
@DatabaseStar
24 күн бұрын
That's true! Others have mentioned it and I wasn't aware of that when I made the video.
that could change work times, greetings
@DatabaseStar
Ай бұрын
Thanks!
Do DuckDB !!
@DatabaseStar
Ай бұрын
Good idea
Awesome
@DatabaseStar
24 күн бұрын
Thanks!
this sounds like a job for DuckDB 🦆🔥💪
@DatabaseStar
24 күн бұрын
Oh good to know!
Very impress
@DatabaseStar
Ай бұрын
Thanks!
Now do the same with duckdb :)
@DatabaseStar
Ай бұрын
Good idea!
wow SQL server and oracle does not exactly cover themselves in glory here, "soptty import to big" does not cut it IMHO. and SQL server needs to step up it's game. I wonder why Mysql is that slow and will mariadb do better?
@DatabaseStar
17 сағат бұрын
True, but I was using the "free" edition of their databases and they are commercial products. I'm not sure why MySQL was slower, perhaps it's not designed for this kind of data. Or maybe the defaults for MySQL are not great for this, and there are settings I can change to improve it.
how to do it in python?
@DatabaseStar
21 күн бұрын
I'm not sure, I don't have a lot of experience in Python. I assume you can write some code to read the CSV file and analyse it directly.
perfect
@DatabaseStar
Ай бұрын
Thanks!
okay than lets try mariadb and than we give them challange under 9 minuts
@DatabaseStar
Ай бұрын
Good idea
due just keep everything the same
@DatabaseStar
Ай бұрын
What do you mean?
Wow! Why is Oracle so fast?
@DatabaseStar
Ай бұрын
I think because it's a more "premium" database and it's built for high-end performance. Or maybe the default settings are better for importing large sets of data compared to the others.
@lucca5101
Ай бұрын
Or because it wasn’t a fair comparison. Both Postgres and SQL Server can directly read external files but you chose to copy the data on all but Oracle.
@yaynative
Ай бұрын
@@lucca5101 Interesting 🤔
@DatabaseStar
Ай бұрын
I couldn't find any documentation on reading a text file directly in Postgres, SQL Server, or MySQL, so I didn't include those in this video. I still included an "import file into table" method for Oracle to see how they would compare as well.