How do you deal with optional fields in documents? do you modify the table schema on the run?
If there's a larg-ish number of optional fields, but each document has only or a few of them, would it create a sparse table with lots of columns? Did you find any problem in these scenarios?
So basically yes. ToroDB creates columns and tables on the fly. All the added columns are nullable, so the ALTER TABLE ADD COLUMN is an almost free operation in PostgreSQL.
Sure, sparse tables are created. This is not a problem since nulls in PostgreSQL are quite cheap (they require no or a few bytes of storage per record).
Even if there is a high cardinality of optional fields, we have not seen in real cases that the number of columns goes beyond a few hundred. And that's perfectly manageable by PostgreSQL :)
There might be some pathological, degraded use cases. But we have found none of them on real datasets.
- I don't see any specs and/or methodology published for any of the benchmarks. I'd like to see some specs for the servers used (especially RAM and what HDD or storage type was used).
- For the 500GB dataset, I assume that it didn't fit in memory, but the 100GB could be "easily" fit in memory. I'd like to see if that's the case, and how does it compare when the dataset is all in memory (I'm anticipating that Mongodb still sucks big time, but it's nice to see a clear apples to apples comparison).
All benchmarks have been done on AWS i2.xlarge instances. They have 4 vCPUs, 30GBs of RAM and 800GB local SSDs.
It is true that we are not doing an apples to apples comparison on the 100GBs set. I think is just the opposite! We are helping MongoDB by giving it three machines (and therefore 90GB RAM!). This is specially true when the benchmark is using indexes because in this case the index can be always on RAM.
MongoDB is very fast when it has to retrieve a single document because, by design, it has an amazing spatial locality (the whole document is usually on the same page). But this feature is a weakness on aggregation queries, as they usually only care about a small subset of the document. ToroDB Stampede change that by storing your data on a relational way. Of course, as you said, MongoDB performance is horrible when it has to fetch documents from disk, but even if the documents are in memory, the same effect is expected (on aggregation queries) when data has to be move to the CPUs caches.
LOL I re-read the article and found the specs. I really read it but somehow managed to skip the paragraph or something :-?
Anyway, my point stands. I'd use a single instance in which the dataset fits in memory, just for completeness. Then, you can compare one mongo with the full dataset to one stampede.
As you said, the aggregated data will still make Mongo suffer, but it will be a better comparison. I still like the 3-shard setup, though. It's also a good reference point.
So far we have benchmarked situations where dataset > RAM or >> RAM. I think it is an interesting point to also analyze the case when dataset < RAM, to see how efficiently both systems manage the caches, query planning etc. Stay tuned and thanks for the suggestion! :)
One of our engineer's recently floated up the idea of using ToroDB Stampede for replacing our MongoDB -> PostgreSQL ETL, since it's merely a set of Python scripts parsing JSON into tables.
Have you seen use cases of Stampede being implemented on existing databases/schemas, rather than an entirely new DB?
One of the most relevant use cases of ToroDB Stampede is precisely what you say, replacing MongoDB to PostgreSQL ETLs, where you have to design the schema, solve data type conflicts, maybe flatten or discard data, etc. And other problems like real-time replication, managing HA, etc. Stampede addresses all these problems.
While most people may want to use an empty, dedicated database for Stampede, it is not required to do so. Stampede will generate the tables under a schema name that matches the collection name. So as long as there are no name conflicts, you can happily have ToroDB-generated tables alongside your own. Needless to say, this gives you the ability to JOIN information from different data sources.
It's so funny to hear sometimes that noSQL is better that SQL databases and see how much more powerful tools can be built using old good relational engines.
Another ToroDB developer here! What you say is one of the things we always repeat: is it really necesary to reinvent the wheel? relational databases have been here since several decades and they have improved from then!
Indeed, MongoDB BI connector v1 was based on PostgreSQL. However, they used 9.4's foreign data wrappers and since those could not push down query clauses, the connector itself introduced a significant performance degradation (on top of MongoDB's current performance difference when compared to Stampede/PostgreSQL).
I've tried that connector and decided against using it.
The problem was you had to define your schema by hand and in mongo sometimes you have legacy data records with different semantics.
In my case it crashed Tableau data import after 45 minutes because it blindly sent data that shul have been NULLIFIED.
Will take a look at the way Stampede does the transformation. It might suit us.
Great! We are looking forward to hear from your experience! Open a ticket on github or email us if you find problems or do you think there is something we can improve!
FYI: myself (and I believe a reasonable number of other software developers) treat the word "Automatic" as a red flag. Magic means you don't understand how something works. Auto(matic) means you don't control something.
Something you don't control and don't understand is dangerous IMO.
I agree sometimes myself think the same way too. In this case, it conveys a non-dangerous, powerful message: rather than having to design your DDL, and have that DDL updated every single time data with a different structure appears in your source stream (MongoDB), that DDL is designed for you. It is automatic because you don't need to do anything. It is "magic" as if ToroDB would be designing the DDL for you, real time, and this is quite disruptive IMHO.
Now, there's no danger: no data is ever lost or mapped incorrectly. Just if you don't like the generated DDL, create some views and done! All data shaped exactly as you want :)
17 comments
[ 2.7 ms ] story [ 48.9 ms ] threadPlease let us know if you have any questions or comments about ToroDB. We will be happy to answer them :)
Enjoy!
If there's a larg-ish number of optional fields, but each document has only or a few of them, would it create a sparse table with lots of columns? Did you find any problem in these scenarios?
Sure, sparse tables are created. This is not a problem since nulls in PostgreSQL are quite cheap (they require no or a few bytes of storage per record).
Even if there is a high cardinality of optional fields, we have not seen in real cases that the number of columns goes beyond a few hundred. And that's perfectly manageable by PostgreSQL :)
There might be some pathological, degraded use cases. But we have found none of them on real datasets.
Stampede looks nice, congrats!
Quick question about the 100x performance claim and the benchmarks in https://www.8kdata.com/blog/announcing-torodb-stampede-1-0-b...:
- I don't see any specs and/or methodology published for any of the benchmarks. I'd like to see some specs for the servers used (especially RAM and what HDD or storage type was used).
- For the 500GB dataset, I assume that it didn't fit in memory, but the 100GB could be "easily" fit in memory. I'd like to see if that's the case, and how does it compare when the dataset is all in memory (I'm anticipating that Mongodb still sucks big time, but it's nice to see a clear apples to apples comparison).
Anyway, kudos for the awesome work!
All benchmarks have been done on AWS i2.xlarge instances. They have 4 vCPUs, 30GBs of RAM and 800GB local SSDs.
It is true that we are not doing an apples to apples comparison on the 100GBs set. I think is just the opposite! We are helping MongoDB by giving it three machines (and therefore 90GB RAM!). This is specially true when the benchmark is using indexes because in this case the index can be always on RAM.
MongoDB is very fast when it has to retrieve a single document because, by design, it has an amazing spatial locality (the whole document is usually on the same page). But this feature is a weakness on aggregation queries, as they usually only care about a small subset of the document. ToroDB Stampede change that by storing your data on a relational way. Of course, as you said, MongoDB performance is horrible when it has to fetch documents from disk, but even if the documents are in memory, the same effect is expected (on aggregation queries) when data has to be move to the CPUs caches.
Anyway, my point stands. I'd use a single instance in which the dataset fits in memory, just for completeness. Then, you can compare one mongo with the full dataset to one stampede.
As you said, the aggregated data will still make Mongo suffer, but it will be a better comparison. I still like the 3-shard setup, though. It's also a good reference point.
Have you seen use cases of Stampede being implemented on existing databases/schemas, rather than an entirely new DB?
One of the most relevant use cases of ToroDB Stampede is precisely what you say, replacing MongoDB to PostgreSQL ETLs, where you have to design the schema, solve data type conflicts, maybe flatten or discard data, etc. And other problems like real-time replication, managing HA, etc. Stampede addresses all these problems.
While most people may want to use an empty, dedicated database for Stampede, it is not required to do so. Stampede will generate the tables under a schema name that matches the collection name. So as long as there are no name conflicts, you can happily have ToroDB-generated tables alongside your own. Needless to say, this gives you the ability to JOIN information from different data sources.
So sure, go and give it a try! :)
Also, this story a year ago story https://www.linkedin.com/pulse/mongodb-32-now-powered-postgr... was simply epic.
In my case it crashed Tableau data import after 45 minutes because it blindly sent data that shul have been NULLIFIED.
Will take a look at the way Stampede does the transformation. It might suit us.
Something you don't control and don't understand is dangerous IMO.
I agree sometimes myself think the same way too. In this case, it conveys a non-dangerous, powerful message: rather than having to design your DDL, and have that DDL updated every single time data with a different structure appears in your source stream (MongoDB), that DDL is designed for you. It is automatic because you don't need to do anything. It is "magic" as if ToroDB would be designing the DDL for you, real time, and this is quite disruptive IMHO.
Now, there's no danger: no data is ever lost or mapped incorrectly. Just if you don't like the generated DDL, create some views and done! All data shaped exactly as you want :)