تدريب Shadowing: The Secret Sauce Behind NoSQL: LSM Tree - تعلم التحدث بالإنجليزية عبر الفيديو

NoSQL databases like Cassandra have exploded in popularity in recent years.

One driver is the insatiable demand to ingest huge amounts of

data from an ever-growing number of sources like mobile application and IoT devices.

The secret sauce behind many of these NoSQL databases is a data structure called Log Structure Merge Tree.

An LSM tree is optimized for fast writes.

To understand how an LSM tree works, let's first take a look at how data is typically stored in a relational database.

A relational database is commonly backed by a data structure called a B-tree.

A B-tree is optimized for reads.

Updating the B-tree is relatively expensive as it involves random I.O and might include updating multiple pages on disk.

This limits how fast a B tree can ingest data.

An LSM tree works differently.

Writes are bashed in memory as they arrive in a structure called a mem table.

A mem table is ordered by object key and is usually implemented as a balanced binary tree.

As a mem table reaches a certain size, it is flushed to disk as an immutable sorted string table.

An ss table stores the key value pairs in a sorted sequence.

These writes are all sequential I.O.

They are fast on any storage media.

The new ss table becomes the most recent segment of the LSM tree.

As more data comes in, more and more of these immutable SS tables are created and added to the LSM tree,

with each one representing a small chronological subset of the incoming changes.

Since SS tables are immutable, an update to an existing object key does not override an existing table.

Instead, a new entry is added to the most recent SS table, which supersedes any entries in the old SS tables for the object key.

Deleting an object requires special handling too, since we cannot mark anything in the SS table as deleted.

To perform a delete, it adds a marker called a tombstone to the most recent SS table for the object key.

When we encounter the tombstone on read, we know the object has been deleted.

And yes, it is a bit unintuitive that a delete takes up extra space.

To serve a read request, we first try to find the key in the mem table, then in the most recent SS table in the LSM tree,

then the next SS table, and so on.

Since an SS table is sorted, the lookup can be done efficiently.

The accumulation of SS tables presents two issues.

As the number of SSTables grows, it will take an increasingly long time to look up a key.

And as the SSTables accumulate, there are more and more outdated entries as keys are updated and tombstones are added.

These take up precious disk space.

To fix these issues, there is a periodic merging and compaction process running in the background

to merge SSTables and discard outdated or deleted values.

This reclaims this space and caps the number of SSTables a read has to look through.

Since the tables are sorted, this merging and compaction process is simple and efficient.

The approach is similar to the one used in the merge phase of the merge sort algorithm.

This, in a nutshell, is how an LSM tree provides fast writes.

Let's recap.

An LSM tree buffers incoming writes in memory.

When the buffer fills up, we sort and flush it to disk as an immutable SS table.

The number of SS tables increases as more buffers get flushed to disk.

This creates a problem for reads, as each read has to search through these SS tables to perform a lookup.

To cap the number of SS tables it has to search through for each read, the LSM tree merges the SSTables and compacts them in the background.

Let's look at the compaction process more closely.

When the SSTables are merged, they are organized into levels.

This is where the tree part of the LSM tree comes into the picture.

There are different strategies to determine when and how the SSTables are merged and compacted.

There are two broad strategies, size-tiered compaction and leveled compaction.

Size tier compaction is optimized for write throughput.

Level compaction is more read optimized.

We won't go into too much detail on how these compaction strategies work.

There are a few key points to remember though.

Compaction keeps the number of SS tables manageable.

The SS tables are organized into levels, with each level gets exponentially larger as SS tables from the level above emerged into it.

Compaction consumes a lot of I.O.

A mistuned compaction could starve the system and slow down both read and write.

Finally, let's visit some common optimizations for the LSM trees in production systems.

There are many optimization strategies that try to provide read performance closer to that of the B tree.

What we explain below are just the common ones, and they are by no means exhaustive.

To look up a key, it performs searches on the SS tables at every level.

Even though search is fast on sorted data, going through all the on-disk SS tables consumes a lot of I.O.

Many systems keep a summary table in memory that contains the min-max range of each disk block of every level.

It allows the system to skip searches on those disk blocks where the key does not fall within the range.

This saves a lot of random I.O.

Another issue that could be potentially expensive is to look up a key that does not exist.

This would require looking through all the eligible blocks at all the levels.

Most systems keep a bloom filter at each level.

A bloom filter is a space-efficient data structure that returns a firm no if a key does not exist, or a probably yes if a key might exist.

This allows the system to skip a level entirely if the key does not exist there, which dramatically reduces the number of random I.O required.

In conclusion, NoSQL databases backed by an LSM tree could be tuned to support a very high write rate.

As with any database, proper tuning is the key.

For an LSM tree, compaction tuning is the most critical.

If you would like to learn more about system design, check out our books and weekly newsletter.

Please subscribe if you learn something new.

Thank you so much and we'll see you next time.

ما هي تقنية التظليل الصوتي؟

التظليل الصوتي (Shadowing) تقنية تعلم لغة مدعومة علمياً، طُورت أصلاً لتدريب المترجمين الفوريين المحترفين. الطريقة بسيطة لكنها قوية: تستمع لصوت إنجليزي أصلي وتكرره فوراً بصوت عالٍ — كظل يتبع المتحدث بتأخير 1-2 ثانية. تُظهر الأبحاث تحسناً كبيراً في دقة النطق والتنغيم والإيقاع وربط الأصوات والاستماع والطلاقة.

تدريب Shadowing: The Secret Sauce Behind NoSQL: LSM Tree - تعلم التحدث بالإنجليزية عبر الفيديو

حول هذا الدرس

ما هي تقنية التظليل الصوتي؟