How RocksDB Works(artem.krylysov.com) |
How RocksDB Works(artem.krylysov.com) |
LSM tree descriptions typically imply or say outright that each layer is laid out linearly, written sequentially, and read sequentally for merging. And that looking up a block within a layer is an O(1) operation, doing random access I/O to that location.
But really, the underlying filesystem is doing a lot of heavy lifting. It's maintaining the illusion of linear allocation by hiding how the large files are fragmented. That sequential writing is mostly sequential, but typically becomes more fragmented in the filesystem layer as the disk gets closer to full, and over time as various uses of the filesystem mean there are fewer large contiguous regions. More fragmented free space makes the allocation algorithms have to do more work, sometimes more I/O, just to allocate space for the LSM tree's "linear" writes.
Lookup of a block inside a layer requires the filesystem to lookup in its extent tree or, with older filesystems, through indirect block lookups. Those are hidden from the LSM tree database, but are not without overhead.
Writing sequentially to a layer generally requires the filesystem to update its free space structures as well as its extent tree or indirect blocks.
Even a simple operation like the LSM tree database deleting a layer file it has finished with, is not necessarily simple and quick at the filesystem layer.
In other words, when analysing performance, filesystems are the unsung heroes underlying some LSM tree databases. Their algorithmic overhead is often not included in the big-O analysis of LSM tree algorithms running over them, but should be, and their behaviour changes as disk space shrinks and over time due to fragmentation.
I think that's vastly under selling what's done to ensure that each block is written linearly, blocks are structured, sized, written and accessed in a way that the filesystem does very little (directio, fadvise, droping caches on writes, etc). I was in total agreement with you, for a long time. The rocksdb devs have put in the work, and tuning rocksdb usually gets faster the less the FS does.
Lately linear reads and writes are not why one is choosing LSM's in a datacenter setting. Access times of even cheap slow ssd's are amazing.
They are used for controlling write amplification with tunable known costs. That is you write fewer hardware blocks to the flash chips with a well tuned rocksdb.
I've worked on my own DB engine that uses a structure similar to LSM (but it's not an LSM tree), where the highest possible performance (millions of TPS) for random-writes mixed with semi-sorted writes mixed with random-reads on current SSDs was the target. There's no need for any data to be allocated sequentially on those, other than just enough aggregation to ensure a sufficiently large block size to reduce IOPS during streaming and merging operations, when IOPS-bound. Indeed it's better to fragment to reuse already filesystem-allocated space where possible - that lowers overhead on the filesystem.
I also agree that a well tuned RocksDB can perform very well, and that the authors have done the work, and that it has methods to reduce avoidable write amplification.
However, the RocksDB applications I've seen haven't use the fancy APIs to get the most out of it. They just used it as a plain k-v store with little understanding of what makes a DB sing, and got not great performance as a result.
What would you consider a well-tuned rocksdb? My understanding is that, due to level-based compaction, there is always a decent amount of write amplification that is unavoidable -- i.e. for one modification to eventually end up in the bottommost level (e.g. L6), it would need to be (re-)written to disk 5 or 6 times. That's quite heavy an amount of write amplification, but maybe my expectations are off?
However, I think you’re making a mistake on a core part of your argument:
> More fragmented free space makes the allocation algorithms have to do more work, sometimes more I/O, just to allocate space for the LSM tree's "linear" writes.
The file system in no way needs to guarantee on-disk contiguity for read or write performance, nor does any online defrag need to happen. Indeed, the whole premise behind LSM trees is to try to optimize around solid state storage. AFAIK if the filesystem can only find 1 MiB blocks it will allocate them at the cost of a larger set of extents (there’s also defrag happening). Typically the filesystems do a fantastic job of defrag too. That’s certainly an important part but I’d say those parts of the filesystem are likely the first things implemented and never/rarely changed (just a hunch - I haven’t actually bothered looking at the Linux changelog).
Also no one is really going to care about performance on an almost full filesystem (kind full like 75% but old so lots of fragments is valid but I doubt it’s actually a problem because of how good filesystems are).
No, that part has been misunderstood so I guess I didn't write it clearly enough.
I'm not saying the filesystem defragments, or does any particular effort to ensure on-disk contiguous storage. I'm saying that as a result of the presence of other data on the filesystem and historic accumulating entropy in layout (sometimes caused by an LSM tree DB!), the filesystem ends up keeping track of appended data discontiguously on disk. In keeping track, the filesystem has to consult its free-space structures to find new free areas, with a heuristic or two to decide whether to allocate a large or small one, write updates to the free-space structures, and keep track of the resulting discontiguous mapping for the file by writing to those structures too. Even when it's contiguous on-disk, versions of those metadata writes are needed for transactional, durable DB writes. They're simpler during bulk non-durable writes of course.
These are the "more work, sometimes more I/O, just to allocate space".
Good filesystems are efficient, but that activity does add overhead (especially when durable transactions are involved) and big-O algorithmic factors (those 1 MiB extents add up), and the picture painted of linear-time operations in LSM papers is inaccurate as a result. I don't think this overhead is necessarily large in practice most of the time. However it's where much of the complexity and corner cases lie, if honestly analysing the performance range and big-O scaling factors.
You make an interesting point about write amplification at the filesystem layer not being accounted for. In addition, classic LSM trees also have significant write amplification (regardless of filesystem or even block device) due to the simple act of writing all data to every layer. This is well known by LSM designers, and there are mitigations, but it's somehow left out of the public image of LSM trees. Classic LSM trees are excellent for random-access writes that need to be gradually sorted, but for some other write patters, are slower (more I/O) than good quality B-trees and some other structures.
> no one is really going to care about performance on an almost full filesystem (kind full like 75% but old so lots of fragments is valid but I doubt it’s actually a problem because of how good filesystems are).
Heh. In practice, every RocksDB or LevelDB I've seen in production is hundreds of GB or several TB on filesystems that have run low on space or even run out, mainly due to the size of the DB :-) They also have thousands or tens of thousands of files in each DB, which they have to open for random-access queries. This is what motivated me to recognise the filesystem can be quite involved.
However, in scenarios that really care about performance, memory-mapped files seem to be able to bypass the I/O stack and optimize performance. RocksDB also has corresponding optimizations: https://github.com/facebook/rocksdb/wiki/PlainTable-Format
files are mapped out in relatively large chunks, especially compaction outputs - there's prealloaction, and usually you will just have a flat file->block conversion without huge trees or anything.
based on performance profiles filesystem doesn't do any heavy lifting, there's not that much fragmentation (and you usually keep some free space for flash GC anyway),
compaction output write is one logical operation on filesystem for tens of megabytes of data.
> filesystems are the unsung heroes underlying some LSM tree databases
meh
A long time ago we had a big MySQL tokudb db and were keen to migrate to myrocks. But myrocks put every table into a single big file, rather than a file per partition.
The partition-per-file is a big deal if you are retaining N days of data in a DB and every night will be dropping some old day. If your DB stores each partition in separate files, the DB can simply delete them. But if your DB stores all the partitions in a single file, then it will end up having to compact your absolutely massive big dataset. It was completely unworkable for us.
Has this changed?
Partitioning data across files (or LSM trees) can be a remarkable win. For data retention policies, as well as for exploiting immutability in different workloads to reduce write amplification.
For example, in TigerBeetle, a DB that provides double-entry financial accounting primitives, our secondary indexes mutate, but half of our ingest volume, all the transactions themselves are immutable, and inserted in chronological order.
We therefore designed our local storage engine as an LSM-forest, putting different key/value types in their own tree, so that mutable data wouldn't compact immutable data. This turns our object tree for primary keys into essentially an append-only log.
I did a lightning talk on this, and a few of our other LSM optimizations, at Jamie Brandon's HYTRADBOI conference last year: https://www.youtube.com/watch?v=yBBpUMR8dHw
RocksDB also allows you to do this, with its concept of column families, if I am not mistaken. However, we wanted more memory efficiency with static memory allocation, deterministic execution and deterministic on disk storage for faster testing (think FoundationDB's simulator but with storage fault injection) and faster recovery (thanks to smaller diffs, with less randomness in the data files being recovered), and also an engine that could solve our storage fault model.
All details in the talk. Or ping me if you have questions.
MyRocks has a collection of files per each column family, and when you drop data it can quickly expunge files that don't contain data for other tables/partitions - and trigger compaction on neighbors, if needed.
RocksDB seems to fit few boxes but there could be much better solution as we don't need deletes/range scans sort of operations.
Any suggestions?
However, most programs do not need to allocate so much memory, so frequently that an allocator become an issue.
It is battle tested. It does one job and does it well.
I have used it in the past as a middleware database taking an average of 2-3k req/sec with over 400 GB of data stored. It works like a charm.
If I had a single reproach to do to it, it would be around the instrumentation. It is not that straightforward to get proper metrics and reporting of the internals.
Both MySQL and ZippyDB are datastores that use RocksDB under the hood, in a slightly different way and with different querying capabilities exposed to the end user. ZippyDB uses it exclusively, but MySQL uses both the traditional InnoDB and RocksDB (MyRocks). TAO is in memory graph database, layer above both of these, and doesn't persist anything by itself - it talks to the database layer (MyRocks).
> Well written article
Thanks!
> It is not Tao it is ZippyDb
I don't work for Meta, so might have made a mistake. There is an old blog post[1] about Tao and there is a recent paper[2] mentioning that the graph database is powered by MyRocks, which runs on RocksDB.
[1]: https://engineering.fb.com/2013/06/25/core-data/tao-the-powe...
(there's not much Cassandra now, though)
> RocksDB runs a dedicated background thread that persists immutable memtables to disk.
They are using "process" to mean "mechanism", something that happens, not a literal OS process. I agree that it's a bit confusing to use the word both ways.
I don't think it's possible except when both key and value are fixed size (which is not the case in the example shown).
There are a number of implementations including Badger (used in dgraph) and a variant that's "RocksDB for large-value use cases" (https://rocksdb.org/blog/2021/05/26/integrated-blob-db.html)
One thing to pay attention to is if your telemetry data is indexed by timestamp (i.e. you're writing to the keyspace in order), the compaction of immutable SSTables layers could be wasteful? Although, the author's nice example of non-overlapping SSTables key ranges suggests there may be minimal write amplification here too.
But really, I would benchmark sqlite first...
Depending on the write pattern, you actually can, because standard LSM-trees write the same data repeatedly, into each layer, and again if recompacting a layer. They make up for it by writing so much sequentially that the gains can outweigh the cost of writing the same data many times, and sorting having a cost no matter how you do it. However, if data is being written in batches in mostly sequential key order, then LSM-trees have a type of write amplification that's worse than efficient B-trees.
However, RocksDB deviates from LSM-trees when writing in bulk (or can be made to), to reduce or avoid this form of write amplification.
The optimal balance depends on the write pattern, but neither standard LSM-trees nor B-trees are consistently better than the other for all write patterns.
Proof needed I think. When last I looked I could get it to just under a gib/s. The disk itself could do 2-3 and ram is 20.
It’s fast but it’s definitely a long long way off from RAM speed. The reason is that the memtable is quite pricey to maintain - you’re having to constantly keep a non trivial amount of data sorted and that sort is expensive.
Maybe I don't understand the problem, but can you not just store it in memory (e.g., with a map from key to current value), update (for instance, increment) it as you go, and whenever you want to take a timeseries value just push the set of current values back to a vector?
Batching sends under the covers to reduce network round trips is all baked in.
This is also one of the things that most existing telemetry clients handle for you ie batching telemetry in memory and shipping it out on an interval, so there's a great deal of existing work you can draw from if not outright copy.
So if the goal is to eventually have the timeseries data merged back to a timeseries DB, and latency isn't too much of a concern then wouldn't batch writing to Kafka (Kinesis, etc) be tolerable?
If you are worried about write amplification then leveled compactions are sub-optimal. I would try the universal compaction.
- https://github.com/facebook/rocksdb/wiki/Universal-Compactio...
If you write (say) 1000 small, contiguous rows in a single transaction, the B-tree wins. The B-tree writes a few full leaf pages and then updates just enough interior pages to point to them. The LSM-tree writes the rows to the first level, then later will write them again to the second level, then to the third layer, then... N times for N levels. The data also has to be read N-1 times.
At (say) 100,000 contiguous rows in a single transaction, the classic LSM-tree is still poor, but a modern LSM-tree with multiple files in the bulk level, with special handling of bulk loads to bypass the levels and write directly to the logical-middle in the bulk level, is similar I/O performance to the B-tree, so it catches up. I think RocksDB has this but you have to request it explicitly, rather than it detecting when to do so automatically. Probably with lower layout discontiguity if it works by writing to files or large contiguous zones. But if we are comparing LSM-tree with modification for bulk loads, we may as well compare with B-tree with a similar modification, which can also use a zone strategy to reduce discontiguity on bulk loads.
There's two different times that it's critical. The first is for write ahead logs. Those are pre-allocated, opened, and ready in order to stop latency spikes.
The second is to control write amplification on ssd writes. If you write 100K bytes. Then issue a hardware flush. Then write another 100K. That's going to be two different writes to the same ssd. Meaning you'll burn it out double fast. That's a huge no no. Flash burn rate is one of the things that LSM's are good at.
But again things have changed since the first times that LSM's became en vouge. Previously it was critically important to write these files out sequentially because it did mean that large parts of the file were laid out sequentially in HHD tracks. That did work. It was important, and it wasn't the filesystem. It was all about issuing huge linear writes that had to be spilled to disk. There were times that the added linear read speed of reading data sequentially was important (read ahead caching, OLAP style queries, etc)
Now the linear reads and writes are structured around SSD block sizes. So 512k or so. Linear reading speed is bottle necked much more on cpu.
>Indeed it's better to fragment to reuse already filesystem-allocated space where possible - that lowers overhead on the filesystem.
I don't understand what filesystem overhead you're referring to here. A read comes in for a key. The SST files are already pre-opened so there's no FS interaction.
Then you walk the index blocks. Almost always those are in cache from just opening the files. (So I'll hand wave a little here but index blocks are just like data blocks with near 100% hit rates)
From there we have walked the index and know that the key we're looking for could be in a specific block. We have the offset in the file to a block.
That block is located on one single logical hardware unit (hdd sector, ssd block, raid stripe size, etc). So the read that finally goes to the fs layer where we translate one read of a index, offset into a read on a single hardware sector, with page cache off, DirectIO is used to. Read ahead is tuned at the application layer not the FS layer. So that will result in a single read operation being sent to the nvme/ssd.
Essentially the SST file is structured such that the reads use extents as a translation layer only of different address spaces the hardware and the file. That translation layer didn't buy LSM's nearly as much as regular file users. And we had to force everything to align so that each read of a key ends up a single hardware sector read.
Contrast that with using the FS more and relying on hardware structure less. I'll assume that walking the index is always in memory as before. So then we need to read some segment of the file. Issue a read. That read goes to the page cache. The page cache has lots of fun with locking and very strange behavior. We're then left with some un-cached data to read. The larger key/value sizes get the more likely it is that you have extra to read. Eg you only need 100k, but that 100k crosses over the ssd block or raid stripe boundary. Now you are waiting on two operations. Those need to fill the page cache, then return the data to waiting buffers.
[0] - https://zonedstorage.io/docs/introduction/zns#overview
I recommend the paper, “Don’t stack your Log on my Log”:
https://www.usenix.org/system/files/conference/inflow14/infl...
Basically, the beautiful thing about log structures is that they work well on “dumb” storage. Paradoxically, when you put logs on top of logs, you can make things worse. This is unintuitive, but it reminds me of the fact that if you tunnel a TCP connection over another TCP connection, the result is something which is much less reliable than an ordinary TCP connection operating over an unreliable network.