Thursday, June 20, 2019

Open Core Summit and OSS glue code

I am optimistic about the Open Core Summit. It can be something that benefits users, startups, the source-available community and the OSS crowd. The summit has many of the important people from the open core community. It is an opportunity for them to collaborate -- form a foundation, reduce open core license proliferation, discuss the next license to be reviewed by OSI and most importantly -- fix the open core marketing approach.

I appreciate that startups put so much time and money into building interesting system software that is shared either as OSS or source-available. I haven't enjoyed much of the marketing about the challenges that cloud creates for VC-funded OSS. I am sure cloud makes it harder but the approach has been too negative with too much snark directed towards OSI and AWS. There is a better way.

Glue code

It is all about the glue code.

Stateful services are a hard problem. It is not trivial to scale MySQL by enabling a small DBA team to support a growing database deployment. I am sure the same is true for other DBMS. This is done with some changes to MySQL and a huge amount of glue code. While the GPL doesn't require sharing of the diffs to MySQL my teams have always been happy to do that. But the glue code is the secret sauce and neither the GPL nor the AGPL require it to be shared. The SSPL would change that, although I am wary of the SSPL given uncertainty about how far down the stack the sharing requirement extends.

While the glue code is the secret sauce I wonder whether it has any value outside of a web-scale company.
  1. The glue code isn't portable as it depends on other systems internal to a web-scale company.
  2. Documentation for internal systems is frequently not good and success depends on the shared knowledge of the current teams. Even with access to the internal dependencies the glue code is unusable without the team that knows how to use it.
Therefore I am more interested in OSS glue code that is useful to a community and less interested in licenses that force the (unusable to me) glue code to be published. The glue code should assume environments open to the public -- public clouds and k8s.

What is the state of OSS glue code for MySQL? MySQL needs it to remain competitive. Much of the glue code is baked into MongoDB so that it is easier to scale MongoDB clusters even if you aren't an Atlas customer.

Thursday, June 13, 2019

Interesting indexing in Rockset and MongoDB

I was lazy today and asked about new indexing features in Rockset and MongoDB. They share a valuable goal which is better indexing for the document data model (think less schema, not schema-less). How do you index documents when you don't know all of the attributes that will be used? MongoDB now supports this via a wildcard index and Rockset via converged indexing.

Wildcard indexing in MongoDB lets you specify that an index should be maintained on all, most or some attributes in a document. By most I mean there are options to exclude attributes from a wildcard index. By some I mean there are options to limit this to attributes that start with certain prefixes. Read the docs for more.

Converged indexing in Rockset indexes all attributes. There are no options to include or exclude attributes. This makes the product easier to use at the cost of more IO for index maintenance and more storage for larger indexes. Note that Rockset uses the RocksDB LSM which reduces the cost of index maintenance and might also use the excellent ZStandard compression.

Wildcard and converged indexes do not support compound indexes. For the document { a:1, b:2 } there will be two index entries: a=1 and b=2. There is no way to get an index entry for (a=1, b=2) or (b=2, a=1). If you want a compound index with MongoDB the existing index features can be used. See below (editorial 1) for compound indexes and Rockset.

Implementation details

This section is an educated guess. I don't know enough MongoDB and Rockset internals to claim this with certainty. I ignore the complexity of support for datatypes. In the ideal world all values can be compared via memcmp.

For a traditional index limited to a specified attribute the index entries are of the form (value, pointer) where pointer points to the row and can be the primary key value or (file name, file offset).

This is more interesting for wildcard/converged indexes. I assume that the attribute name is the leading field in each index entry so that the entry is of the form (attribute name, value, pointer). The common way to use such an index is to have an equality predicate on attribute name which is satisfied when the index is queried with predicates like attributeName relOp value. Examples of such predicates are a=2, a>2 and a<=2.

A smart person (Dr Pavlo) mentioned the use of skip scan for these indexes. That could be used to query the index and find documents with any attribute equal to a specific value. That is a less likely use case but still interesting.

Wildcard/converged indexes aren't free. Putting the attribute name in every index entry makes index entries larger and consume more space in memory and on storage. Block compression reduces some of this overhead. Index prefix compression in WiredTiger and RocksDB also helps but at the cost of more CPU overhead.

Storage differences

Up to now I have been describing the search index. In this section I will describe the document storage.

MongoDB stores the document via the storage engine which will soon be WiredTiger only although I hope MongoRocks returns. I assume that WiredTiger with MongoDB is row-wise so that each document is (usually) a contiguous sequence of bytes on some disk page.

Rockset stores each document twice -- row-wise and column-wise. Alas, this gets complicated. The row-wise format is not the traditional approach with one thing in the storage engine per document. Instead there is one thing per attribute per document. This is similar to the CockroachDB approach. I prefer to still call this row-wise given that attributes from a document will be co-located in the LSM SSTs. I am also grateful for the many great blog posts from CockroachDB that I can reference.

With two copies of each document in the base storage there is more storage overhead. Fortunately that overhead is reduced courtesy of the write efficiency and compression friendliness of an LSM.

The Rockset blog post does a great job of explaining this with pictures. I do a worse job here without pictures. For the document { pk:1, a:7, b:3 } when the primary key is pk then the keys for row-wise are R.1.a and R.1.b and for column-wise are C.a.1 and C.b.1. The row-wise format clusters all attributes for a given document. The column-wise format clusters all values across documents for a given attribute. The row-wise format is efficient when most attributes for a document must be retrieved. The column-wise format is efficient for analytics when a given attribute across all documents must be retrieved.

Editorial 1

I interpret the MongoDB docs to mean that when a query uses a wildcard index it cannot use any other index and the wildcard index will only be used for a predicate on a single attribute. I expect that to limit the utility of wildcard indexes. I also expect MongoDB to fix that given how fast they reduce their tech debt. The limitations are listed below. The 1st won't be fixed. The 2nd and 3rd can be fixed.
  1. Compound wildcard indexes are not supported
  2. MongoDB cannot use a non-wildcard index to satisfy one part of a query predicate and a wildcard index to satisfy another.
  3. MongoDB cannot use one wildcard index to satisfy one part of a query predicate and another wildcard index to satisfy another.
I assume that Rockset can combine indexes during query evaluation given their focus on analytics. Thanks to the Rockset team I learned it supports index intersection. It also supports composite indexes via field mappings (functional indexes).

Editorial 2

An open question is whether an LSM can do clever things to support analytics. There has been some work to show the compression benefit from using column-wise storage within a RocksDB SST for the larger levels of the RocksDB LSM. Alas, the key RocksDB workloads have been OLTP. With the arrival of Rockset there is more reason to reconsider this work. There can be benefits in the compression ratio and reduced overhead during query processing. Vertica showed that it was useful to combine a write-optimized store for recent writes with a read-optimized store for older writes. An LSM already structures levels by write recency. Perhaps it is time to make the larger levels read-optimized especially when column-wise data is inserted to the LSM.

Update - read paper years ago then forgot that Kudu combines LSM + columnar.

Editorial 3

The previous section is mostly about being clever when storing column-wise data in an LSM to get better compression and use less CPU during query evaluation. This section is about being clever when storing the search index. 

The search index is likely to have many entries for some values a given attribute. Can an LSM be enhanced to take advantage of that for analytics workloads? In other storage engines there are two approaches -- bitmap indexes and RID-lists. Adapting these for an LSM is non-trivial but not impossible. It is likely that such an adaptation would only be done for the larger levels of the LSM tree.

Friday, May 17, 2019

index+log: implementations

My take on index+log systems like WiscKey is that they are neither better nor worse than an LSM - it all depends on your workload. But I am certain that we know much more about an LSM than about the index+log systems. Hopefully that changes over time as some of them are thriving.

The first index+log system that I read about was Berkeley DB Java Edition. The design paper is worth reading. Since then there have been a few more implementations and papers that I describe here. This list is probably incomplete: Bitcask, ForestDB, WiscKey, HashKV, TitanDB and RocksDB BlobDB.

At this point the systems that are getting into production, TitanDB and BadgerDB, use an LSM for the index. I wonder if an index structure that supports update-in-place would be better especially when the index must be cached because I expect the CPU read-amp for an LSM to be about 2X larger than for a b-tree and a b-tree supports update-in-place which makes it easier to relocate values during GC.

While I like index+log systems I think that papers and marketing tend to overstate LSM write-amp. For production RocksDB I usually see write-amp between 10 and 20. I expect that index+log could achieve something closer to 5. This paper from CMU explains one reason why per-level write-amp in an LSM is less than the per-level fanout (less than 10). Write skew is another reason.

The systems

Bitcask was part of the Riak effort.
  • The index is an in-memory hash table. The index isn't durable and the entire log has to be scanned to rebuild it on startup -- whether or not this was after a crash or a clean shutdown. The slow startup is a problem.
  • The value log is circular and GC copies live values from the tail to the head of the log. Liveness is determined by an index search. 

ForestDB was a finalist in the SIGMOD 2011 student programming contest. Eventually the project and creator moved to Couchbase. It is worth reading about here and on the github page. I published blog posts that compare ForestDB and RocksDB: 1, 2, 3 and 4. Google finds more interesting things to read.
  • The index is a space-efficient trie.
  • The value log might have log segments. GC copies live values to the head of the log. Liveness is determined by an index search.

WiscKey is described as an LSM with key-value separation and made popular the term key-value separation. I put it in the index+log family of index structures.
  • The index is an LSM. There is no redo log for the index as it can be recovered from the head of the value log.
  • Kudos for many references to amplification factors. The paper uses bytes read for read-amp. I prefer to consider both IO and CPU for read-amp with key comparisons for CPU and storage reads for IO.
  • It doesn't mention that it has more cache-amp than an LSM, but few papers mention that problem. Shrinking the LSM by keeping large values separate doesn't make the index parts of it (filter and index blocks) easier to cache as they are already separate from the data blocks. There is more to cache with index+log as I describe here.
  • It claims to overcome the (worst-case) problem of one storage IO per KV pair on a range scan by fetching in parallel. Assuming the storage device has enough spare IO this might hide the problem but it doesn't fix it. With many workloads there isn't spare IO and extra IO for reads also means extra CPU for decompression.
  • The value log is circular and single-threaded GC copies live values to the head of the log. Liveness is determined by an index search. I assume that multi-threaded GC is feasible.
  • The paper isn't clear about the total write-amp that might occur from both the first write to the value log and GC that follows.
  • Compression isn't explained.

BadgerDB is a golang implementation, and much more, of the WiscKey paper.
  • It has many features and many production use cases. This is impressive. 
  • GC is scheduled by the user. Based on Options.NumCompactors I assume it can be multi-threaded.
  • The docs state that the LSM can be served from RAM because the values are elsewhere. That is true but I don't consider it a feature. It must be in RAM to avoid IO from liveness queries done by GC. An LSM isn't a monolithic thing. There are index blocks, data blocks and filter blocks and most of the LSM, data blocks from the max level, don't have to be cached. 
  • There is extra work on reads to find values that have been moved by GC. See the comments about BadgerDB here.

HashKV is an interesting paper that avoids index queries during GC.
  • Hash-based data grouping distributes KV pairs by hash into one of N logs. GC is probably done by scanning a log twice -- once to get the keys and the second time to relocate the live values. A value is live when the most recent key is not a tombstone. A value might be live when needed for a snapshot. GC doesn't do index searches so the index doesn't have to be cached to make GC efficient but you might want to cache it to avoid doing index IO on queries -- and this index is still much larger than the block index for an LSM.
  • Hotness awareness copies cold values to a cold segment to avoid repeated GC for a value that doesn't get updated or deleted. A header for the value is kept in the non-cold log.
  • Small values are stored inline in the LSM.
  • I am curious if more log groups means more write-amp. See my comment about fsync in a previous post.
  • I am curious whether managing the hash buckets is easy. The goal is to make sure that keys for a segment group fit in memory. The range and number of buckets must change over time. Does this have anything in common with linear and extendible hashing?

TitanDB is part of TiDB and TiDB is thriving.
  • A WAL is used for new writes. This might make it easier to compress data on the first write to the value logs.
  • GC appears to do index searches to determine liveness.
  • Compression is per-record. I hope this does per-block in the future.
  • It might let the user tune between space-amp and write-amp via discardable_ratio.
  • This is compatible with most of the RocksDB API.k

RocksDB BlobDB is an extension to RocksDB that uses log segments for large values and stores small values in the LSM. GC copies live values and liveness is determined by an index search.

Future work

Future work for index+log systems includes:
  • Determine whether a b-tree is better than an LSM for the index structure
  • Determine whether the HashKV solution is the best way to avoid liveness queries during GC.
  • If an LSM is used for the index structure determine efficient ways to support relocating values during GC without extra overheads and complexity during read processing.
  • Determine whether clever things can be done during GC.
    • Block compression is easier than on first write.
    • Hot/cold value separation is possible (see HashKV). This is an example of generational GC even if we rarely mention that for index structures.
    • Values in a log segment can be ordered by key rather than by time of write during GC. GC can also merge multiple ordered segments to create longer sorted runs. I wonder if it is then possible to use block indexes (key+pointer per block rather than per row) to reduce cache-amp for such log segments.

Thursday, May 16, 2019

index+log: open issues

This post is about open issues for the index+log family of index structures. See past posts here and here for more details on index+log. The success of LSM has lead to several solutions that use an LSM for the index. I am curious if that is a good choice when implementing index+log from scratch. It is a great choice when starting with an LSM and then adding index+log.

Open Issues

The open issues for index+log include:
  • Is GC multi-threaded?
  • Does value log GC require index queries?
  • Must the index be cached?
  • Is block compression possible?
  • With an LSM as the index how is relocating an entry in the value log supported?

Is GC multi-threaded?
  1. I hope so.
  2. It hasn't been in some of the index+log papers/systems.
Does value log GC require index queries?
  1. Yes, in most of the index+log papers/systems this is required to determine whether a value is live when scanning the value log during GC
  2. HashKV proposed a solution that doesn't require such queries. It is a wonderful paper but I think there are more things to figure out. The general idea is to hash KV pairs into buckets such that all keys for a bucket will fit in memory during GC. GC then reads the value log segments for a bucket twice -- first to get the keys, second to copy the live KV pairs into a new log segment. I wonder if managing the hash bucket ranges has things in common with linear and extendible hashing.
Must the index be cached?
  1. The index for index+log might be 10X larger than for an LSM because an LSM uses a block index while index+log uses a row index. A block index has a key+pointer per block and a row index has that per row. An LSM only needs a block index because rows are clustered by key.
  2. To do at most one IO per point query for a database larger than memory the index must be cached for index+log to avoid index IO as the IO is spent reading the value log. If the <= 1 IO / query constraint is then cache-amp is larger for index+log compared to LSM because the index+log index is larger (see #1).
  3. If value log GC queries the index to determine whether each value is live then this query shouldn't do IO or GC will be too slow and inefficient. This is another reason for caching the index+log index.
  4. If the index must be cached then maybe an LSM isn't the best choice. Consider an index structure optimized for a main-memory DBMS.
Is block compression possible?
  1. I hope so but this hasn't been explained in some of the index+log papers/systems.
  2. Per-record compression can be done instead of block compression. That will have a lower compression rate but less CPU overhead when decompressing on read.
  3. It might be hard to do block compression the first time a KV pair is written to a value log. One option is to write to a redo log until enough data is available to do block compression and then write to the value log. Another option is to defer block compression until KV pairs are rewritten during GC.
When the index is an LSM what happens when values are moved?
  1. This is an issue that I learned about via CockroachDB. I should have figured it out long ago but mistakes happen.
  2. The LSM tree is read in order -- top to bottom with leveled compaction and left to right with tiered compaction. This guarantees that the correct result is returned with respect to visibility. If the first entry for a key is a tombstone then the search can stop (ignoring snapshot reads).
  3. Value log GC moves live KV pairs to new log segments. To find the KV pair after the move either the index entry must be updated or the index entry must have a logical value log key and then another index is needed to map the logical value log key to a physical value log (filename, offset). 
  4. Updating the LSM index entry to reference the new value log location (filename, offset) can be done by inserting a new KV pair into the LSM but that either breaks read consistency semantics or complicates read processing. It would break LSM read processing because inserting back into the LSM implies this is a new write, but it just a move for GC. Something other than an LSM that supports update-in-place makes this easier.
  5. Details on using a logical value log key are explained in a CockroachDB github issue.

Wednesday, May 15, 2019

Index+log, v2

I put most index structures into one of three categories -- page-based, LSM or index+log. My focus is on databases larger than memory and I might be ignoring categories used for main memory DBMS. Over the past decade index+log has gotten more attention and this is my second attempt at explaining it.

My definition of index+log is simple -- data is appended to a log on each write, index entries point into the log and GC scans the log to copy live values and discard others. The log can be one file or many log segments. GC might have to search the index to determine whether a value is live. The value written to the log might include the key.

Bitcask was the first index+log system that I noticed but I assume there are earlier examples and just found one -- Berkeley DB Java Edition. While WiscKey made popular the term key-value separation and is presented as an LSM variant, I put it in the index+log category. Other interesting index+log systems include RocksDB BlobDBTitanDBForestDB and Faster. While many of the tree-based solutions use an LSM for the index that is not required by index+log and an LSM is not used by Berkeley DB Java Edition or ForestDB.

For an index+log solution that must cache all of the index and query the index during GC I suspect that an LSM is not the best choice for the index. Although if you already have an LSM (see RocksDB BlobDB) then I get it.

The summary of my LSM vs index+log comparison is:
  1. index+log has less write-amp but more space-amp
  2. index+log has much more cache-amp, maybe 10X more
  3. index+log might have more deferred CPU write-amp
  4. I have yet to validate my claims by running benchmarks with index+log implementations.
Note that #1 is a feature, #2 is not a feature and for #3 it depends. The key point is that the cost of faster writes from index+log is more cache-amp (more memory) and more IO for range queries. In production with RocksDB I frequently see write-amp=15 with space-amp=1.1. I assume that could be reduced with index+log to write-amp ~=5 and space-amp ~= 1.3. It might be possible to avoid or reduce the impact of #2 and #3 in future index+log implementations.

Amplification Factors

I am speculating on read, write, space and cache amplification for index+log because I have little hands on experience with index+log implementations. Another reason for speculation is that index+log allows for different structures for the index (b-tree, LSM, etc) which affects some of the estimates.

The amplification factors are:
  • cache - the cache-amp for index+log is likely to be much larger than for an LSM. To achieve at most one IO per point query index+log might need 10X (or more) memory versus an LSM. Clustering values in key order doesn't just make range queries faster, it also means an LSM only needs block indexes in memory (key+pointer per block) while index+log needs a key+pointer in memory for every KV pair in the database. When there are 10 KV pairs per block then this is 10X more memory. Even when the database has hot and cold keys they are likely to be interleaved on the same index leaf pages --  so all of those leaf pages must be in memory to avoid doing two IOs (one for the leaf page, one for the value) on a point query.
    • There is additional data that an LSM and b-tree need in memory to satisfy the one IO per query constraint and I described that in previous posts (mostly things from the non leaf/max level).
    • It might be valid to debate whether my one IO per point query constraint is valid, but this blog post is already long.
    • Another reason for the index to be cached is to avoid doing IO during GC when index searches are done to determine whether KV pairs are live.
  • space - I ignore space-amp for the index and focus on the log because the log is usually larger. With index+log the user can trade between write-amp and space-amp. With the variable pct_full representing the percentage of live data in the database (a value between 1 and 100) then:
    • space-amp = 100 / pct_full
    • write-amp = 100 / (100 - pct_full)
    • Just like with an LSM, previously written KV pairs are rewritten during GC with the index+log approach. Fortunately this is done less often with index+log. 
    • I assumed that block compression is used but that is harder to implement for index+log. The WiscKey paper doesn't describe a solution and the HashKV paper suggests using per-record compression, which will have a lower compression rate versus block as used by an LSM. I assume block compression can be done for index+log but it isn't trivial.
    • To explain the estimates when pct_full=X assume that all log segments have X% live data (yes, this is naive). When GC is done on a log segment X% is copied leaving (100-X)% free space in the newly written log segment. So in total 100% of a log segment is written for each (100 - pct_full)% of new data, which is the formula above.
    • Thus with pct_full=90 then space-amp is 1.1 while write-amp is 10. Comparing these with a leveled LSM the space-amp is similar while the write-amp is slightly better than what I see in production. To get a write-amp that is significantly better the cost is more space-amp. For example with pct-full=75 then write-amp=4 and space-amp=1.33. 
  • read (CPU) - see here for range seek/next. The summary is that when an LSM is used for index+log then the costs are similar to an LSM. When a b-tree is used for index+log then the costs are smaller.
  • read (IO) - see here for range seek/next. In the cache-amp estimate I assume that the index is cached so the only IO to be done is for the log. Therefore the index structure (LSM vs b-tree) doesn't matter.
    • point query - the IO read-amp is <= 1 because the log is not cached.
    • range seek - range seek doesn't do IO when the index is cached
    • range next - this is much larger for index+log than for an LSM because it might do one IO per call to range next because rows are not clustered by key in the log. When data is compressed then there also is the CPU overhead for decompression per KV pair.
  • write - by write I assume update (read-modify-write) rather than a blind write.
    • immediate CPU - the cost of an index search. See the section on read CPU for point queries above.
    • immediate IO - the cost of an optional redo log write for the index structure and then writing the (value) log. Note that the minimum size of a write done by the storage device might be 4kb even if the data written is much smaller. Doing an fsync per 128 byte value might have a write-amp of 32 if that write is really forced to storage and doesn't just linger in a capacitor backed write cache.
    • deferred CPU - the deferred CPU write-amp is the cost of index searches done during GC, unless the HashKV approach is used. With pct_full=75, write-amp=4 and space-amp=1.33 then GC is done ~4 times for each key and the deferred CPU cost is 4 index searches. When live KV pairs are copied by GC then there is also the CPU and IO overhead from updating the index entry to point to the new location in the log.
    • deferred IO - this is determined by the percentage of live data in the database. With pct_full=75 it is 4.

Thursday, May 9, 2019

CRUM conjecture - read, write, space and cache amplification

The RUM Conjecture asserts that an index structure can't be optimal for all of read, write and space. I will ignore whether optimal is about performance or efficiency (faster is better vs efficient-er is better). I want to use CRUM in place of RUM where C stands for database cache.

The C in CRUM is the amount of memory per key-value pair (or row) the DBMS needs so that either a point query or the first row from a range query can be retrieved with at most X storage reads. The C can also be reported as the minimal database : memory ratio to achieve at most X storage reads per point query.

My points here are:
  • There are 4 amplification factors - read, write, space and cache
  • CRUM is for comparing index structure efficiency and performance
  • Read and write amplification have CPU and IO parts
  • Write amplification has immediate and deferred parts
Many X is faster than Y papers and articles neglect to quantify the tradeoffs made in pursuit of performance. I hope that changes and we develop better methods for quantifying the tradeoffs (a short rant on defining better).

Amplification factors (RUM -> CRUM) are used to compare index structures. Values for the factors are measured for real workloads and estimated for hypothetical ones. The comparison is the most valuable part. Knowing that the deferred CPU write-amp on inserts for a b-tree is 30 is not that useful. Knowing that it is 3X or 0.5X the value for an LSM is useful.

Workload matters. For estimates of amplification I usually assume uniform distribution because this simplifies the estimate. But there is much skew in production workloads and that impact can be measured to complement the estimates.

Read Amplification

This post is an overview for read-amp. This post explains it in detail for an LSM. There are two parts to read-amp -- CPU and IO. Thus for each of the three basic operations (point query, range seek, range next) there are 2 values for read-amp: CPU and IO. I have yet to consider deferred read-amp and by read-amp I mean immediate read-amp.

Metrics for CPU read-amp include CPU time, bytes/pages read and key comparisons. I use key comparisons when predicting performance and CPU time when running tests. I have not used bytes/pages read. While key comparisons are a useful metric they ignore other CPU overheads including hash table search, bloom filter search, page read and page decompress.

Metrics for IO read-amp include bytes read and pages read. I use pages read for disk and bytes read for SSD because disks are IOPs limited for small reads. IO read-amp implies extra CPU read-amp when the database is compressed. Decompressing pages after storage reads can use a lot of CPU with fast storage devices and even more with zlib but you should be using zstd.

With estimates for hypothetical workloads I assume there is a cache benefit as explained in the Cache Amplification section. This is likely to mean that comparisons assume a different amount of memory for index structures that have more or less cache-amp. For real tests I mostly run with database >> memory but don't attempt to use the least memory that satisfies the cache-amp X reads constraint.

Write Amplification

This post is an overview for write-amp. This post explains it in detail for an LSM. Write-amp has two dimensions: CPU vs IO, immediate vs deferred. For each operation (insert, delete, update) there are 4 values for write-amp: immediate CPU, deferred CPU, immediate IO and deferred IO.

The immediate write-amp occurs during the write. The deferred write-amp occurs after the write completes and includes writing back dirty pages in a b-tree and compaction in an LSM.

Possible metrics for CPU write-amp include bytes written, pages written, key comparisons and pages/bytes (de)compressed. Bytes and (in-memory) pages written are useful metrics for in-memory DBMS but my focus is on databases >> memory.

Possible metrics for IO write-amp include bytes written and pages written. These can be estimated for hypothetical workloads and measured for real ones. The choice between bytes or pages written might depend on whether disk or SSD is used as one is limited by ops/s and the other by transfer rate. If you use iostat to measure this then figure out whether Linux still counts bytes written as bytes trimmed.

Examples of deferred and immediate write-amp:
  • The InnoDB change buffer is deferred IO and CPU. Checking the change buffer and applying changes is deferred CPU. The deferred IO is from reading pages from storage to apply changes.
  • For a b-tree: page writeback for a b-tree is deferred IO, compression and creating the page checksum are deferred CPU, finding the in-memory copy of a page is immediate CPU, reading the page on a cache miss is immediate IO.
  • An LSM insert has immediate/deferred IO/CPU. 
    • Immediate CPU - key comparisons for memtable insert
    • Immediate IO - redo log write
    • Deferred IO - reading uncached pages for input SSTs and writing output SSTs during compaction
    • Deferred CPU - decompression, compression and key comparisons while merging input SSTs into output SSTs during compaction. Note that compaction does a merge, not a sort or merge+sort.

Space Amplification

Space-amp is the size of the database files versus the size of the data, or the ratio of the physical to logical database size. An estimate for the logical size is the size of the uncompressed database dump with some adjustment if secondary indexes are used. The space-amp is reduced by compression. It is increased by fragmentation in a b-tree and uncompacted data in an LSM.

It is best to measure this after the DBMS has reached a steady state to include the impact of fragmentation and uncompacted data.

Cache Amplification

I briefly described cache-amp in this post. The cache-amp describes memory efficiency. It represents the minimal database : memory ratio such that a point query requires at most X storage reads. A DBMS with cache-amp=10 (C=10) needs 10 times more memory than one with C=100 to satisfy the at most X reads constraint.

It can be more complicated to consider cache-amp for range seek and range next because processing them is more complicated for an LSM or index+log algorithm. Therefore I usually limit this to point queries.

For a few years I limited this to X=1 (at most 1 storage read). But it will be interesting to consider X=2 or 3. With X=1:
  • For a b-tree all but the leaf level must be in cache
  • For an LSM the cache must include all bloom filter and index blocks, all data blocks but the max level
  • For an index+log approach it depends (wait for another blog post)

Other posts

Related posts by me on this topic include:

Friday, April 12, 2019

A research paper on Optane performance

I just read Basic Performance Measurements of the Intel Optane DC Persistent Memory Module published by the NVSL at UCSD. It is worth reading. I appreciate the rigor in testing and the separation of the summary (first 10 pages) from the many details. This is too incomplete to be a review of the paper. It is really a collection of my comments.


  • When using the device in cached mode where RAM is the cache the cache block size is 4kb. I assume that a cache miss does 16 256-byte reads from the Optane device before returning to the user.
  • The paper doesn't explain the endurance for the device. The word "endurance" doesn't occur in the paper. I read elsewhere that Optane might provide ~60 DWPD. Update - I assume that endurance isn't mentioned because the vendor has yet to disclose that info.
  • The paper states that the Optane DIMM uses a protocol that supports variable response time but doesn't explain how much it varies. How does response time variance in Optane compare to a NAND-flash SSD where the stalls can be bad?
  • The Optane DIMM does 256 byte reads and writes. I wonder if that prevents 4kb page writes from being atomic when this is used for a filesystem assuming copy-on-write isn't done internally, as it might be for Nova.
  • There is wear-leveling. I am not sure whether that has a name yet. I saw one blog post that called it the XTL to match the FTL used by NAND flash. I am also curious about the latency impact from doing lookups on the XTL to determine locations for 256 byte blocks. The XTL is cached in RAM and a 256g device needs ~4g of RAM assuming each 256 byte block uses 4 bytes in the XTL.
  • Nova does much better than XFS and ext4 on Optane. Nova is a research filesystem from NVSL that exploits new features in Optane.
  • They modified RocksDB to make the memtable persistent and avoid the need for a WAL. It will be interesting to learn whether that turns out to be useful.

Requests I have for the next Optane performance paper:
  • For mixed and concurrent workloads include response time latencies -- average+variance or a histogram. This paper reports the average latency for single-threaded read-only and write-only. For mixed+concurrent workloads this paper reports average throughput which combines read and write performance. It is hard to determine whether reads or writes degrade more from concurrency and a mixed workload. 
  • For any workload include information about response time variation whether that is variance or a histogram
  • Provide numbers to accompany graphs because some of the graphs are hard to understand without numbers when the lines converge in one part and diverge in another because the range for the y-axis is large. Figure 18 is one example. 

Tuesday, January 22, 2019

Less "mark" in MySQL benchmarking

My goal for the year is more time learning math and less time running MySQL benchmarks. I haven't done serious benchmarks for more than 12 months. It was a great experience but I want to learn new things. MySQL 8.0.14 has been released with fixes for a serious bug I found via the insert benchmark. I won't confirm whether it has been fixed. I hope someone else does.

My tests and methodology are described in posts for sysbench, linkbench and the insert benchmark.  I hope the upstream distros (MySQL, MariaDB, Percona) repeat my tests and methodology and I am happy to answer questions about that. I even have inscrutable shell scripts that make it easy to run the tests. Despite being a lousy example of how to use Bash, they are portable enough to run on my home and work hardware.

Monday, January 21, 2019

Optimal configurations for an LSM and more

I have been trying to solve the problem of finding an optimal LSM configuration for a given workload. The real problem is larger than that, which is to find the right index structure and the right configuration for a given workload. But my focus is RocksDB so I will start by solving for an LSM.

This link is to slides that summarizes my effort. I have expressed the problem to be solved using differentiable functions to express the cost that is to be minimized. The cost functions have a mix of real and integer valued parameters for which values must be determine to minimize the cost. I have yet to solve the functions, but I am making progress and learning more math. This might be a constrained optimization problem and Lagrange Multipliers might be useful. The slides are from a talk I am about to present at the MongoDB office in Sydney where several WiredTiger developers are based. I appreciate that Henrik Ingo set this up.

My work has things in common with the excellent work by Harvard DASlab lead by Stratos Idreos. I have years of production experience on my side, they have many smart and ambitious people on their side. There will be progress. I look forward to more results from their Data Calculator effort. And I have learned a lot from the Monkey and Dostoevsky papers by Niv Dayan et al.

Sunday, January 20, 2019

Bugs in Windows 10 parental controls

I use Windows 10 parental controls with my two children. Sometimes I am surprised at the bugs I encounter, but I can't rant too much because of glass houses and stones. My old favorite was that a hard reset before the time limit reached zero allowed my clever child to get more time. Apparently Microsoft takes storage efficiency very seriously and didn't want to waste a disk write and/or fsync on persisting the usage counter every few minutes. I haven't tried to reproduce this recently but never heard back after filing a bug report.

Now I have a new favorite bug. I am 5 hours behind their timezone and granted another hour to my daughter. It is 4pm here and 9pm there. The landing page after granting the time tells me my child can use the computer until 5pm (my timezone).  Child tries to login and immediately encounters the timeout dialog. Apparently timezones are a hard problem. But less screen time is a good thing.

Tuesday, January 15, 2019

Geek code for LSM trees

This is a link to slides from my 5-minute talk at the CIDR 2019 Gong Show. The slides are a brief overview of the geek code for LSM trees. If you click on the settings icon in the slide show you can view the speaker notes which have links to blog posts that have more details. I also pasted the links below. Given time I might add to this post, but most of the content is in my past blog posts. Regardless I think there is more to be discovered about performant, efficient and manageable LSM trees.

The key points are there are more compaction algorithms to discover, we need to make it easier to describe them and compaction is a property of a level, not of the LSM tree.

Links to posts with more details:

Thursday, January 10, 2019

LSM math: fixing mistakes in my last post

My last post explained the number of levels in an LSM that minimizes write amplification using 3 different estimates for the per-level write-amp. Assuming the per-level growth factor is w then the 3 estimates were approximately w, w+1 and w-1 and named LWA-1, LWA-2 and LWA-3 in the post.

I realized there was a mistake in that post for the analysis of LWA-3. The problem is that the per-level write-amp must be >= 1 (and really should be > 1) but the value of w-1 is <= 1 when the per-level growth factor is <= 2. By allowing the per-level write-amp to be < 1 it easy to incorrectly show that a huge number of levels reduces write-amp as I do for curve #3 in this graph. While I don't claim that (w-1) or (w-1)/2 can't be a useful estimate for per-level write-amp in some cases, it must be used with care.

Explaining LWA-3

The next challenge is to explain how LWA-3 is derived. That comes from equation 12 on page 9 of the Dostoevsky paper. Start with the (T-1)/(K+1) term and with K=1 then this is (T-1)/2. T in the paper is the per-level growth factor so this is the same as (w-1)/2. The paper mentions that this is derived using an arithmetic series but does not show the work. I show my work but was not able to reproduce that result.

Assume that the per-level growth factor is w, all-to-all compaction is used and the LSM tree has at least 3 levels. When full L1 has size 1, L2 has size w and L3 has size w*w. There are four derivations below - v1, v2, v3, v4. The results are either w/2 or (w+1)/2 which doesn't match (w-1)/2 from the paper. Fortunately, my previous post shows how to minimize total write-amp assuming the per-level write-amp is w/2 or (w+1)/2. I will contact the author to figure out what I am missing.

The analysis below is for merges from L1 to L2, but it holds for merges from Ln to Ln+1. I think that v1 and v2 are correct and their estimate for per-level write-amp is (w+1)/2. As explained below I don't think that v3 or v4 are correct, their estimate for per-level write-amp is w/2.

I have yet to explain how to get (w-1)/2.


Assume that merges are triggered from Ln to Ln+1 when a level is full -- L1 has size 1, L2 has size w, L3 has size w*w. A level is empty immediately after it is merged into the next level. So L2 gets full, then is merged into L3 and becomes empty, then slowly gets larger as L1 is merged into it w times. The per-level write-amp from this is (w+1)/2.

* merges into L2 write output of size 1, 2, ..., w
* then L2 is full
* sum of that sequence -> w*(w+1)/2
* average value is sum/w -> (w+1)/2

1) Moving data of size 1 from L1 to L2 writes (w+1)/2 on average
2) Therefore per-level write-amp for L1 -> L2 is (w+1)/2

Note that per-level write-amp is (avg merge output to Ln / size of Ln-1)
* avg merge output to L2 is (w+1)/2
* size of Ln-1 is 1


Assume that merges are triggered from Ln to Ln+1 when a level is almost full -- L1 has size 1 * (w-1)/w, L2 has size w * (w-1)/w, L3 has size (w*w) * (w-1)/w. The trigger conditions can be reduced to L1 has size (w-1)/w, L2 has size (w-1) and L3 has size w*(w-1).

This assumes that w merges are done from L1 to L2 for L2 to go from empty to full. Each merge adds data of size (w-1)/w because L1:L2 merge is triggered when L1 has that much data. Thus L2 has size (w-1) after w merges into it at which point L2:L3 merge can be done. The per-level write-amp from this is the same as it was for v1.

* merges into L2 write output of size (w-1)/w * [1, 2, ..., w]
* then L2 is full
* sum of that sequence -> (w-1)/w * w*(w+1)/2 = (w-1)(w+1)/2
* average value is sum/w -> (w-1)(w+1)/(2*w)

As from v1, per-level write-amp is (avg merge output to Ln / size of Ln-1)

* avg merge output to L2 = (w-1)(w+1)/(2*w)
* size of L1 = (w-1)/w

start with: ( (w-1)(w+1)/(2*w) ) / ( (w-1)/w )
simplify to: (w+1)/2


Merges are triggered the same as for v1 but I assume that only w-1 merges are done from Ln to Ln+1 rather than w. Ln+1 won't be full at the end of that, for example L2 would have size w-1 rather than the expected size w. But I was curious about the math. The per-level write-amp is w/2.

* merges into L2 write output of size 1, 2, ..., w-1
* sum of that sequence -> (w-1)*w/2
* average value is sum/(w-1) -> w/2

1) Moving data of size 1 from L1 to L2 writes w/2 on average
2) Therefore per-level write-amp for L1 -> L2 is w/2


Merges are triggered the same as for v2. But as with v3, only w-1 merges are done into a level. Again I don't think this is correct because a level won't have enough data to trigger compaction at that point. The per-level write-amp here is the same as for v3.

* merges into L2 write output of size (w-1)/w * [1, 2, ..., w-1]
* sum of that sequence -> (w-1)/w * (w-1)*w/2 = (w-1)(w-1)/2
* average value is sum/(w-1) -> (w-1)/2

As from v1, per-level write-amp is (avg merge output to Ln / size of Ln-1)

* avg merge output to L2 = (w-1)/2
* size of L1 = (w-1)/w

start with: ( (w-1)/2 ) / ( (w-1)/w )
simplify to: w/2

Wednesday, January 9, 2019

LSM math: revisiting the number of levels that minimizes write amplification

I previously used math to explain the number of levels that minimizes write amplification for an LSM tree with leveled compaction. My answer was one of ceil(ln(T)) or floor(ln(T)) assuming the LSM tree has total fanout = T where T is size(database) / size(memtable).

Then I heard from a coworker that the real answer is less than floor(ln(T)). Then I heard from Niv Dayan, first author of the Dostoevsky paper, that the real answer is larger than ceil(ln(T)) and the optimal per-level growth factor is ~2 rather than ~e.

All of our answers are correct. We have different answers because we use different functions to estimate the per-level write-amp. The graph of the functions for total write-amp using the different cost functions is here and you can see that the knee in the curve occurs at a different x value for two of the curves and the third curve doesn't appear to have a minimum.

While working on this I learned to love the Lambert W function. But I wonder whether I made the math below for LWA-2 harder than necessary. I am happy to be corrected. I appreciate the excellent advice on Quora: here, here and here. The online graphing calculator Desmos is another great resource.


I use differentiable functions to express the total write-amp as a function of the number of levels, then determine the value (number of levels) at which the first derivative is zero as that might be the global minimum. Constants, variables and functions below include:
  • T - total fanout, = size(database) / size(memtable)
  • n - number of levels in the LSM tree
  • LWA, LWA-x - function for the per-level write-amp
  • TWA, TWA-x - function for the total write-amp, = n * LWA
  • w - per-level growth factor, = T^(1/n) for all levels to minimize write-amp
The function for total write-amp has the form: TWA = n * LWA where n is the number of levels and LWA is the per-level write-amp. LWA is a function of T and n. The goal is determine the value of n at which TWA is minimized. While n must be an integer the math here doesn't enforce that and the result should be rounded up or down to an integer. T is a constant as I assume a given value for total fanout. Here I use T=1024.

I wrote above that the 3 different answers came from using 3 different estimates for the per-level write-amp and I label these LWA-1, LWA-2 and LWA-3. When w is the per-level growth factor then the per-level write-amp functions are:
  • LWA-1 = w -- I used this to find that the best n = ceil(ln(T)) or floor(ln(T))
  • LWA-2 = w + 1 -- with this the best n is less than that found with LWA-1
  • LWA-3 = (w - 1) / 2 -- with this the best n is greater than that found with LWA-1
I can also state the per-level write-amp functions directly with T and n. I didn't above to make it easier to see the differences.
  • LWA-1 = T^(1/n)
  • LWA-2 = T^(1/n) + 1
  • LWA-3 = (T^(1/n) - 1) / 2
Explaining LWA

First I explain LWA-1 and LWA-2. Compacting 1 SST from Ln to Ln+1 requires merging 1 SST from Ln with ~w SSTs from Ln+1 where w=10 by default with RocksDB. The output will be between w and w+1 SSTs. If the output is closer to w then LWA-1 is correct. If the output is closer to w+1 then LWA-2 is correct. This paper explains why the per level write-amp is likely to be less than w. Were I to use f*w where f < 1 for LWA-1 then the math still holds. Maybe that is a future blog post.

LWA-3 assumes that all-to-all compaction is used rather than some-to-some. I explain the difference here. RocksDB/LevelDB leveled uses some-to-some but all-to-all is interesting. With all-to-all when compaction from Ln to Ln+1 finishes then Ln is empty and slowly gets full after each merge into it. Assume the per-level growth factor is w and Ln-1, Ln and Ln+1 are full at sizes 1, w and w*w. Then Ln becomes full after w merges from Ln-1 and those write output of size 1, 2, ..., w-1, w. The sum of the first w integers is w(w+1)/2. Divide this by w to get the averge -- (w+1)/2. However above LWA-3 is (w-1)/2 not (w+1)/2. I will explain that in another blog post. Note that in LWA-3 the numerator, w-1, is more interesting than the denominator, 2. Dividing by any constant doesn't change where the minimum occurs assuming there is a minimum and that is visible on this graph that shows the impact of dividing by 2 on the total write-amp.

Read on to understand the impact of using w-1, w or w+1 as the function for per-level write-amp. The difference might be more significant than you expect. It surprised me.

Minimizing TWA

This graph shows the total write-amp for LWA-1, LWA-2 and LWA-3. I call the total write-amp TWA-1, TWA-2 and TWA-3. Two of the curves, for TWA-1 and TWA-2, appear to have a minimum. One occurs for x between 4 and 6, the other for x between 6 and 8. The third curve, for TWA-3, doesn't appear to have a minimum and is decreasing as x (number of levels) grows.

The next graph uses the first derivative for the total write-amp functions, so it is for TWA-1', TWA-2' and TWA-3'. A global minimum for TWA-x can occur when TWA-x' = 0 and from the graph TWA-1'=0 when x=6.931 and TWA-2'=0 when x=5.422 which matches the estimate from the previous paragraph. From the graph it appears that TWA-3' approaches zero as x gets large but is never equal to zero.

The next step is to use math to confirm what is visible on the graphs.

Min write-amp for LWA-1

See my previous post where I show that n = ln(T) minimizes total write-amp if n isn't limited to an integer and then the per-level growth factor is e. Since the number of levels must be an integer then one of ceil(ln(T)) or floor(ln(T)) minimized total write-amp.

Min write-amp for LWA-2

I can reuse some of the math from my previous post. But this one is harder to solve.

# wa is the total write-amp
# n is the number of levels

# t is the total fanout
wa = n * ( t^(1/n) + 1 )
wa = n*t^(1/n) + n

# the difference between this and the previous post is '+1'

wa' = t^(1/n) + n * ln(t) * t^(1/n) * (-1) * (1/n^2) + 1
wa' = t^(1/n) - (1/n) * ln(t) * t^(1/n) + 1

At this point the difference between this and the previous post is '+1'. But wait this starts to get interesting.

# critical point for this occurs when wa' = 0
t^(1/n) - (1/n) * ln(t) * t^(1/n) + 1 = 0

# multiply by t^(-1/n)
1 - (1/n) * ln(t) + t^(-1/n) = 0

# move some terms to RHS
t^(-1/n) = (1/n) ln(t) - 1

# use ln on LHS and RHS to get rid of '^(1/n)'
ln ( t^(-1/n) ) = ln( (1/n) * ln(t) - 1 )
(-1/n) ln(t) =  ln( (1/n) * ln(t) - 1

I got stuck here but eventually made progress.

# let a = (1/n) ln(t) and rewrite
-a = ln(a - 1)

# let x=a-1, a=x+1 and rewrite
-(x+1) = ln(x)

# do e^LHS = e^RHS
e^-(x+1) = e^ln(x)
e^-x * e^-1 = x

# multiply LHS and RHS by e^x
e^-1 = e^x * x

# e^-1 -> (1/e)
(1/e)  =  e^x * x

At last I can use Lambert W function!

# Given: e^x * x = K, then x = W(K)
x = W(e^-1) ~= 0.27846

# because a=x+1
a ~= 1.27846

# a = (1/n) ln(t) -> n = (1/a) ln(t), t=1024
n = 1/1.27846 * ln(1024)

# The value for n that minimizes total write-amp
# from the graph I claimed that n=5.422. this is close
n = 5.4217

Min write-amp for LWA-3

Update-1 - I think I made a few mistakes here. So you can stop reading until update-2 arrives.

Update-2 - this post explains my mistake and uses math to estimate that per-level write-amp = (w+1)/2 when all-to-all compaction is used. I am still unable to derive (w-1)/2.

I started to work on this without paying attention to the curve for LWA-3'. From the graph it appears to converge to 0 but is always less than 0, TWA-3 is decreasing as x, number of levels, gets large. Therefore make the number of levels as large as possible, 2M or 2B, to minimize total write-amp as visible in this graph.

But more levels in the LSM tree comes at a cost -- more read-amp. And the reduction in write-amp is small when the number of levels increases from 20 to 200 to 2000 to 2M. Again, this is visible in the graph. Besides, if you really want less write-amp then use tiered compaction rather than leveled with too many levels.

The other consideration is the minimal per-level growth factor that should be allowed. If the min per-level growth factor is 2. Then then that occurs when the number of levels, n, is:

# assume total fanout is 1024

2^n = 1024
log2(2^n) = log2(1024)
n = log2(1024) = 10

Alas the total fanout isn't always a power of 2. Given that the number of levels must be an integer then the goal is to use the smallest number of levels such that the per-level growth factor >= 2. Therefore when x isn't limited to an integer there is no answer -- just make x as large as possible (1M, 1B, etc) in which case the per-level growth factor converges to 1 but is always greater than 1.

The above can be repeated where the constraint is either the max number of levels or a different value for the min per-level growth factor (either <2 or >2). Regardless, if LWA-3 is the cost function then total write-amp is minimized by using as many levels as possible subject to these constraints.

Below is some math for LWA-3 and LWA-3'.

# wa is the total write-amp
# n is the number of levels

# t is the total fanoutwa = n * ( t^(1/n) - 1 ) / 2
wa = (n*t^(1/n) - n ) / 2

# the big difference between this and the previous post is '+1'

wa' = [ t^(1/n) + n * ln(t) * t^(1/n) * (-1) * (1/n^2) - 1 ] / 2
wa' = [ t^(1/n) - (1/n) * ln(t) * t^(1/n) - 1 ] / 2

# determine when wa' = 0
[ t^(1/n) - (1/n) * ln(t) * t^(1/n) - 1 ] / 2 = 0

# multiply LHS and RHS by 2
t^(1/n) - (1/n) * ln(t) * t^(1/n) - 1 = 0# multiply LHS and RHS by t^(-1/n)
1 - (1/n) * ln(t) - t^(-1/n) = 0

# move last term to RHS
1 - (1/n) * ln(t) = t^(-1/n)

# probably a good idea to stop here
# LHS is likely to be <0 so can't use ln(LHS) = ln(RHS)

Monday, January 7, 2019

Define "better"

Welcome to my first rant of 2019, although I have written about this before. While I enjoy benchmarketing from a distance it is not much fun to be in the middle of it. The RocksDB project has been successful and thus becomes the base case for products and research claiming that something else is better. While I have no doubt that other things can be better I am wary about the definition of better.

There are at least 3 ways to define better when evaluating database performance. The first, faster is better, ignores efficiency, the last two do not. I'd rather not ignore efficiency. The marginal return of X more QPS eventually becomes zero while the benefit of using less hardware is usually greater than zero.
  1. Optimize for throughput and ignore efficiency (faster is better)
  2. Get good enough performance and then optimize for efficiency
  3. Get good enough efficiency and then optimize for throughput
Call to action

I forgot to include this before publishing. Whether #1, #2 or #3 is followed I hope that more performance results include details on the HW consumed to create that performance. How much memory and disk space were used? What was the CPU utilization? How many bytes were read from and written to storage? How much random IO was used? I try to report both absolute and relative values where relative values are normalized by the transaction rate.

Thursday, January 3, 2019

Review of LSM-based Storage Techniques: A Survey

Chen Luo and Mike Carey published a wonderful survey of research on LSM algorithms. They know about LSM because the AsterixDB project includes an LSM. They did a great job explaining the LSM space, telling a coherent story and summarizing relevant papers. Reading this paper was a good use of my time and I found a few more papers to read in their references.

I have read a few papers, including TRIAD, with ideas on reducing write-amp for the smaller levels of the LSM tree. I think this could be done for RocksDB by merging and remerging immutable memtables -- this is similar in spirit to subcompactions for the L0. With a large immutable memtable there would be one less level in the LSM tree. This is an alternative to having an L0, and maybe an L1, that are not made durable. In all cases the cost is a longer MTTR because WAL replay must be done. In all cases there is an assumption that the non-durable levels (large immutable memtables or L0/L1) are in memory.

This is a small complaint from me that I have made in the past. The paper states that an LSM eliminates random IO when making things durable. I prefer to claim that it reduces random IO. With leveled compaction each step merges N (~11) SSTs to generate one steam of output. So for each step there is likely a need to seek when reading the ~11 input streams and writing the output stream. Then compaction steps usually run concurrently when the ingest rate is high so there are more seeks. Then the WAL must be written -- one more stream and a chance for more seeks. Finally user queries are likely to read from storage causing even more seeks.  Fortunately, there will be fewer seeks per insert/update/delete compared to a B-Tree.

The paper has a short history of compaction describing pure-tiered and pure-leveled. But these are rarely used in practice. The original LSM paper implemented pure-leveled. LevelDB and RocksDB use a hybrid approach with tiered for the L0 followed by leveled for the remaining levels. Pure-tiered was introduced by the Stepped Merge paper. Using tiered for all levels has a large space-amplification, much larger than 1, because the max level is tiered and that is too much wasted space for many workloads. Tiered in RocksDB and other popular LSM engines can be configured to use leveled compaction into the max level to get a space-amp less than 2, ignoring transient space-amp during compaction into the max level. Pure-tiered was a great choice for Stepped Merge because that was a cache for bulk-loading a data warehouse rather than a full copy of the database. While I think that RocksDB leveled and RocksDB tiered are examples of tiered+leveled, I don't want to rename them.

I appreciate that the paper makes clear that trade-offs must be considered when evaluating benchmarks. Many things can support higher write rates than RocksDB with leveled compaction, including RocksDB with tiered compaction. But that comes at a cost in memory, read and/or space amplification. Some papers could do a better job of documenting those costs.

The cost analysis in section 2.3 is limited to IO costs. I look forward to coverage of CPU costs in future LSM research. The read penalty for an LSM compared to a B-Tree is usually worse for CPU than for IO. The paper uses partitioned and non-partitioned where I use all-to-all and some-to-some to explain the compaction approaches. RocksDB implements some-to-some for leveled and all-to-all for tiered. The paper does a nice job explaining why the per-level write-amp should be less for all-to-all than some-to-some, ignoring write skew. Note that in production the per-level write-amp is almost always less than the per-level growth factor and this paper from Hyeontaek Lim explains why.

For the read IO costs, the paper counts logical IOs rather than physical IOs. Logical IOs are easier to estimate because caches mean that many logical IOs don't cause a physical IO and smaller levels in the LSM tree are usually in cache. There are two ways to consider the cost for a range query -- long vs short range queries or the cost of range seek vs range next. The paper uses the first, I use the second. Both are useful.

I appreciate that the author noticed this. I realize there is pressure to market research and I am not offering to try and reproduce benchmark results, but I have been skeptical about some of the comparisons I see where the base case is InnoDB or RocksDB.
These improvements have mainly been evaluated against a default (untuned) configuration of LevelDB or RocksDB, which use the leveling merge policy with size ratio 10. It is not clear how these improvements would compare against a well-tuned LSM-tree.
The discussion in 3.3.1 on pipelining compaction is interesting but RocksDB already does pipelining. With buffered IO there is support for async read-ahead and async write-behind. Note that the read and write phases can also be CPU-heavy if the cost for decompression on read and compression on write are included, even when the wonderful zstd and lz4 algorithms are used.

A few more comments:
  • RocksDB has limited support for fractional cascading (from SST to SST). See 3.4.2.
  • With key-value separation, GC could merge log segments to generate longer ordered log segments over time. This would reduce the range read penalty. See 3.4.2.
  • LHAM might be the first time-series optimized compaction strategy. See 3.5.
  • Non-unique secondary index maintenance is already read-free in MyRocks. It has a copy of the row prior to index maintenance, because SQL semantics or because this was an insert. Write-optimized SQL engines can add support for read-free change statements in some cases but that usually means SQL semantics (like modified row count) will be broken. See 3.7.2.
  • MyRocks already collects statistics during compaction. See 3.7.3.