What DB do you use at your end? | Bitcointalksearch.org

TransaDox

full member

Activity: 219

Merit: 102

Quote from: btc_enigma on December 09, 2016, 01:58:26 AM

mysql is slow in terms of read performance

Why not use leveldb as used by core. It has very good read speeds. Also you don't require concurrent writes. Only need to write a block that comes in every 10 minutes. Most of the times are reads. I doubt any other RDBMS can match the performance of leveldb

Because LevelDB is not ACID compliant and prone (read, renowned for) corrupting the DB. This is why we are being told to back up the chain, make copies etc. Bitcoin should be using an ACID compliant DB and the only one that gets close to the performance of a Key/Value DB for sequential access is SQLite. Plus, one has other niceties like live backup and in memory attached DBs.

clickerz

hero member

Activity: 1414

Merit: 505

Backed.Finance

Quote from: fikihafana on December 17, 2016, 11:58:31 AM

I would use mongodb to maximize performance and scalability

MongoDB is quite good and in fact it is gaining more followers also. It is also now incorporated in MEAN Stack development and it is becoming popular. I like MongoDb for its scalability. So far my also with performance is common database being used especially on web development.

piotr_n

legendary

Activity: 2053

Merit: 1354

aka tonikt

It is not about the language/API of the queries, but about the speed of accessing the records.

In SQL database just parsing the SQL query takes too much time.
But modern NoSQL databases (like MongoDB) are also not good here, because they are designed for complex queries and fancy data structures.
This all comes at a cost of performance and memory usage.

What you need for bitcoin is a simple [binary_key] -> [binary_record] database.
LevelDB is the most common one I know that can be used for that.
But IMHO even LevelDB is far away from optimal for bitcoin's UTXO set.

I'd say that if you want your bitcoin node to perform really well, you ought to make your own DB engine.
Depending how much RAM you can spare, I'd lean towards keeping all the data inside the memory.
Write data to the disk only to make the memory changes persistent and read from the disk only when loading the DB data while booting your node.
This requires more than 2 GB of RAM for the current UTXO-set, but gives the optimal performance.

That's for UTXO database - which is the tricky one here.
As for the blocks, you might just as well use a pure file system - no additional database engine is really necessary here.

PremiumCodeX

hero member

Activity: 1204

Merit: 531

Metaverse 👾 Cyberweapons

Quote from: piotr_n on December 17, 2016, 12:07:25 PM

You guys have no idea what you are talking about.

At this moment UTXO database holds over 45 millions records, from over 16 millions transactions.

What you need is an instant access to any of these records - otherwise verifying of blocks and transactions will be too slow.

No SQL database (or mongodb) can give you such a performance.
These kind of DBs are meant for complex queries, but they would have killed your bitcoin node had you tried to use them for handling the UTXO set.

Are you suggesting using a No-SQL DB? I have not used a NoSQL DB yet. Is it any better than SQL DBs when handling BTC transaction data? It could be a good idea since BTC is often looked as an experiment, why not to use a new technology with it.

piotr_n

legendary

Activity: 2053

Merit: 1354

aka tonikt

You guys have no idea what you are talking about.

At this moment UTXO database holds over 45 millions records, from over 16 millions transactions.

What you need is an instant access to any of these records - otherwise verifying of blocks and transactions will be too slow.

No SQL database (or mongodb) can give you such a performance.
These kind of DBs are meant for complex queries, but they would have killed your bitcoin node had you tried to use them for handling the UTXO set.

fikihafana

sr. member

Activity: 462

Merit: 250

I would use mongodb to maximize performance and scalability

TransaDox

full member

Activity: 219

Merit: 102

Quote from: CIYAM on December 05, 2016, 08:49:13 AM

That certainly is a surprise to me (thanks for the link and I deleted my own post to avoid any further confusion). Thinking about it more I guess it makes sense that in order to do a "rollback" it would need to be able to restore deleted data (as the size of a transaction could be huge it has to write all the recovery information and keep that at least until the commit has been completed and all final tx data has been flushed to disk).

But can you force rollbacks to occur after a successful commit (and especially after a checkpoint)?

Certainly I'm not familiar with any standard SQL commands for doing such a thing (and I wouldn't count nested transactions if supported as really being a solution) so I still don't see how you could use a traditional RDMBS to achieve a blockchain "re-org" (which is the main point that I was perhaps poorly trying to make).

Assuming that most RDBMS engines do work the same way regarding logging then perhaps it would at least be potentially possible to add some method of tagging the DB state at a certain "block height" and therefore later be able to "rewind" the DB back to the state it was at that specified tag (although I don't think that there is any such standard method).

Rollback is a bit of a red herring for an immutable block chain. Just store everything, good or bad, and adjust the query.

SQLite also has the ability to treat multiple, separate database files as a single database at run-time (max 64). One could create separate DB files for working DB and historical but I don't even see the need for that.

BTW. SQLite can insert 1,000,000 records in under 10 secs and wipes the floor with My/MSSQL and Postgres in terms of performance. What it doesn't have is authentication so the others are a better choice for web servers.

tardi

newbie

Activity: 35

Merit: 0

I would insert the outputs from bitcoin into a mysql database and then query it.

btc_enigma

hero member

Activity: 688

Merit: 565

mysql is slow in terms of read performance

Why not use leveldb as used by core. It has very good read speeds. Also you don't require concurrent writes. Only need to write a block that comes in every 10 minutes. Most of the times are reads. I doubt any other RDBMS can match the performance of leveldb

uanode

newbie

Activity: 62

Merit: 0

We are using PostgreSQL for these tasks whose performance and capabilities are endless.

PremiumCodeX

hero member

Activity: 1204

Merit: 531

Metaverse 👾 Cyberweapons

I am glad because I have not closed this discussion yet lol. Thank you Rick Storm for enlightening me about this matter! I think, you could do this "tagging" with triggers. You could define the block height as a custom condition in a trigger and in the trigger set a savepoint where data is stored too. I too am not sure about how you could "tell the DBMS" to do that so I would be thankful if someone attached some example in any SQL.

CIYAM

legendary

Activity: 1890

Merit: 1072

Ian Knowles - CIYAM Lead Developer

That certainly is a surprise to me (thanks for the link and I deleted my own post to avoid any further confusion). Thinking about it more I guess it makes sense that in order to do a "rollback" it would need to be able to restore deleted data (as the size of a transaction could be huge it has to write all the recovery information and keep that at least until the commit has been completed and all final tx data has been flushed to disk).

But can you force rollbacks to occur after a successful commit (and especially after a checkpoint)?

Certainly I'm not familiar with any standard SQL commands for doing such a thing (and I wouldn't count nested transactions if supported as really being a solution) so I still don't see how you could use a traditional RDMBS to achieve a blockchain "re-org" (which is the main point that I was perhaps poorly trying to make).

Assuming that most RDBMS engines do work the same way regarding logging then perhaps it would at least be potentially possible to add some method of tagging the DB state at a certain "block height" and therefore later be able to "rewind" the DB back to the state it was at that specified tag (although I don't think that there is any such standard method).

Rick Storm

newbie

Activity: 17

Merit: 0

Quote from: ?? on ??

But what you don't have in the log is the content of the record XXX which therefore cannot be restored unless you actually restore an older backup (this is the meaning of D in ACID transaction which is what all serious RDBMS do - i.e. once deleted always deleted).
...
and it should be clear that they could not do this simply because the information has not been preserved in the transaction log

This doesn't sound right. Afaik all major databases store the deleted record in the transaction log, e.g. SQL Server:

https://sqlfascination.com/2010/02/03/how-do-you-decode-a-simple-entry-in-the-transaction-log-part-1/

PremiumCodeX

hero member

Activity: 1204

Merit: 531

Metaverse 👾 Cyberweapons

Thank you for suggestion! I too read very positive opinions on NEXUS, seems worthy to try out next.

@CIYAM, I find your description a bit difficult to imagine. Maybe because I have studied for my finals the whole day and my brain is tired Tongue

Have you got a model or - even better a - prototype of such a DB to see?

borris123

hero member

Activity: 728

Merit: 500

Nexus has created their own DB 10x faster than googles.

https://twitter.com/mirraxFTC/status/798217450739220481

https://bitcointalksearch.org/topic/nexus-pure-sha3-cpugpu-npos-15-active-innovations-more-to-come-657601

piotr_n

legendary

Activity: 2053

Merit: 1354

aka tonikt

Quote from: PremiumCodeX on November 17, 2016, 04:13:45 PM

If you use an additional DB at your end (other than core's leveldb) to store blockchain data, what DB do you use?

I use my own solution, based on a hashmap (acting as the index) holding pointers to the records.

It makes accessing the db very fast, but needs much more RAM comparing to leveldb or berkeley.

Apart from the hashmap index and the records, there is of course also a code for the disk operations - to keep the files in sync with the memory.

It's simple and rather archaic, but has been proven to work great for UTXO-db purposes - I would not swap it for any other db engine I know.
Although I'm also quite interested in other solutions, which I don't know - there must be some more optimal ways of doing it.

https://github.com/piotrnar/gocoin/tree/master/lib/qdb

CIYAM

legendary

Activity: 1890

Merit: 1072

Ian Knowles - CIYAM Lead Developer

Quote from: TransaDox on November 28, 2016, 08:50:03 PM

SQLite enables one to create "Savepoints" which are like GOTO statements for rollback.

An interesting feature although from reading the documentation I think that is just for use with nested transactions.

The concept of a blockchain re-org though is not the same thing as you need to "rollback" transactions that have already been committed (something rather alien to how RDBMS transactions operate).

TransaDox

full member

Activity: 219

Merit: 102

Quote from: CIYAM on November 27, 2016, 07:31:34 AM

A "true rollback" (which as I said isn't being used by an major RBDMS that I am aware of) would actually be much faster than any other method (i.e. you either are going to have to issue update queries to act as though things are being unwound or restore an earlier backup which you'd then have to perform a partial log restore from).

Understand also that even if deletes involve appending the data to the log you would still be able to truncate the log for "checkpoints" (at which point any of that extra wasted space is recovered).

SQLite enables one to create "Savepoints" which are like GOTO statements for rollback.

BuySomeBitcoins

sr. member

Activity: 434

Merit: 253

The choice of the database depends on your needs, mostly I work on market places and cybersecurity [fraud] so I use MongoDB - Oracle

CIYAM

legendary

Activity: 1890

Merit: 1072

Ian Knowles - CIYAM Lead Developer

Quote from: PremiumCodeX on November 27, 2016, 06:50:31 AM

It depends on the use-case then because in some DBs there are frequent delete operations. But I agree that most DBs are not like that so the possibility is worth to take into consideration.

For blockchains (whose very purpose is to be a non-mutating structure) the "deletes" are only re-orgs (which although not that uncommon are generally not very large delete operations as most re-orgs are only one or two blocks deep) and depending upon how things are structured (such as if the structure of the chain itself is separated from the content of the blocks) would not need to involve much data.

Quote from: PremiumCodeX on November 27, 2016, 06:50:31 AM

The question is how much it would increase the necessary space and the issue of SRP because even though rollbacks are infrequent operations, when they need them they usually need them ASAP.

A "true rollback" (which as I said isn't being used by an major RBDMS that I am aware of) would actually be much faster than any other method (i.e. you either are going to have to issue update queries to act as though things are being unwound or restore an earlier backup which you'd then have to perform a partial log restore from).

Understand also that even if deletes involve appending the data to the log you would still be able to truncate the log for "checkpoints" (at which point any of that extra wasted space is recovered).

Topic: What DB do you use at your end? (Read 3654 times)