Neo: Use RocksDB?

[lock9]

Should we use RocksDB instead of LevelDB?

[shargon]

RocksDb https://github.com/facebook/rocksdb/wiki/Features-Not-in-LevelDB

[lock9]

@shargon is there any security or consistency problem we could have by using that?

[erikzhang]

Is there a .NET Standard library for RocksDB?

[shargon]

This nuget is for dotnet https://github.com/warrenfalk/rocksdb-sharp

[shargon]

@shargon is there any security or consistency problem we could have by using that?

No, and maybe is better because now we are using rocksdb in mac, this unify all the storage layer

[Tommo-L]

There is a inconsistency bug before leveldb 1.22, https://github.com/google/leveldb/issues/320.
But it has been fixed in rocksdb long long ago.

[shargon]

it is already decided? @neo-project/core

[erikzhang]

Not yet.

[igormcoelho]

For me, if it's safe enough to use natively on nugets, it's better than current one (easier to use, test and maintain). So I agree.

Too many code changes are expected on this DB "migration", or just few? During these changes, it would be nice to leave things more flexible, so that on the future, it becomes even easier to change that (mainly to support memory-only databases).

[lightszero]

Rocksdb is developed base on leveldb,so it is very easy to use it for NEO.

[shargon]

Take into account that rocksDb only works on x64 computers

[vncoelho]

@jsolman, what is your opinion?

[jsolman]

There is nothing wrong with using LevelDB as long as using the release that fixes the compaction bug mentioned earlier.

[lock9]

It looks like the majority of people are either in favor of this or at least "not against".
I like this because it will allow me to run neo-cli in MacOS

[Qiao-Jin]

Some leveldb bugs might probably occur in NEO like https://github.com/google/leveldb/issues/320, but don't exist in ricksdb.
Rocksdb also provides better write efficiency & functionalities.
Moreover, rocksdb's API is very similiar to leveldb, which makes it convinient to transfer from leveldb to rocksdb.
I have tried the following experiment in Neo system:

1. Directly convert .ldb files to .sst ones by modifying file extension.
2. Use rocksdb's library in Native.cs funtions instead of leveldb's.

Node can be started without any problem.

P.s.
Rocksdb's API: https://github.com/facebook/rocksdb/blob/master/include/rocksdb/c.h

Migrating from LevelDB to RocksDB: https://rocksdb.org/blog/2015/01/16/migrating-from-leveldb-to-rocksdb-2.html

Benchmarking LevelDB vs. RocksDB vs. HyperLevelDB vs. LMDB Performance for InfluxDB: https://www.influxdata.com/blog/benchmarking-leveldb-vs-rocksdb-vs-hyperleveldb-vs-lmdb-performance-for-influxdb/

[Qiao-Jin]

Using rocksdb might also be a solution for this issue as I commented: https://github.com/neo-project/neo/issues/597#issuecomment-519766426

[lock9]

@neo-project/core we don't have a voting system in place yet, but it seems that we have a large majority of people that vote for RocksDB. Can I move this to 'ready to implement'?

[shargon]

Rocksdb works only on x64, maybe we should be able to work with both of them, and switch by configuration

[lock9]

@shargon maybe this is a good option to keep compatibility.
@neo-project/core Can we move this forward? I think we have sufficient votes and a solution with backward compatibility

[lock9]

@neo-project/core Can we move this to ready-to-implement? This solves a lot of problems. Thanks.

[shargon]

I think that we should use column families intead of prefixes, what do you think @neo-project/core
https://github.com/facebook/rocksdb/wiki/Column-Families#targetText=Each%20key-value%20pair%20in,to%20logically%20partition%20the%20database.

[vncoelho]

@shargon, I think that it was not yet decided to migrate to RocksDB.

@jsolman and @erikzhang did not show appreciation for this migration.
I also did not really investigated this enough. I search for some benchmarks and did not get the real feeling that this would be better.

@igormcoelho, can you please double check this with someone specialized in DBs?

[shargon]

The last comment is from 7 days ago

@neo-project/core Can we move this to ready-to-implement? This solves a lot of problems. Thanks.

And there are no more opinions until my pull request 😕

[vncoelho]

I understand, @shargon, but I think that the police is not "silence is consent"...ahauehau

I do not know, I think that this needs a better discussion. Unless you are sure that RocksDB is going to have a better performance and other needed features.

[vncoelho]

@shargon, I really think that we need to move in a direction in which we make a chance for it to become more flexible, not really swap LevelDB for RockDB, but just create an option for switching it.

[shargon]

Now it's switchable https://github.com/neo-project/neo-cli/pull/451

[vncoelho]

Even with Column Families? Nice job. It looks like it will be more readable.
Edited: I saw it is now isolated.

[shargon]

I think that column families could help to save space, 1 byte per entry, and also, will save a lot of concatenations

[erikzhang]

Can we use RocksDb on x64 and LevelDb on x86? Or should we just abandon x86 since there is no 32 bits computer on the world any more.

[shargon]

Yes, we can change by config, or made LevelDB mandatory on x86.

Configuration: default -> RocksDB

if the system is x86 -> use LevelDB mandatory

[shargon]

Some benchmarks
https://www.influxdata.com/blog/benchmarking-leveldb-vs-rocksdb-vs-hyperleveldb-vs-lmdb-performance-for-influxdb/

[vncoelho]

I saw this benchmark before, it is almost the only one I really found in favor of RocksDB, I read other discussions in favor of leveldb but without tests.

But this benchmark is not very well experimentally designed, in my point of view, I usually do not consider this as a final answer, that is why I am still not convinced that it is really more efficient for our case. However, I have faith it will.

It will be good that we can switch by any of them, then, anyone can just which one to use, @shargon.
In addition, that idea of just running on local memory will also make us design something each time more flexible.
I believe that your implementation will be good and open such doors.

[erikzhang]

I have another option: FASTER. https://github.com/microsoft/FASTER

FASTER is a concurrent key-value store + cache that is designed for point lookups and heavy updates. FASTER supports data larger than memory, by leveraging fast external storage. It also supports consistent recovery using a new checkpointing technique that lets applications trade-off performance for commit latency.

Detailed analysis of C# FASTER performance: https://github.com/Microsoft/FASTER/wiki/Performance-of-FASTER-in-C%23

FASTER is implemented in two languages: C# and C++.

The performance of the C# and C++ versions of FASTER are very similar.

@neo-project/core Let's consider it.

[shargon]

I considered it in the past and i think that @PeterLinX discarded it after his research. But I will review it again, maybe now is better!

[lightszero]

how about try to build a x86 version rocksdb？