Let’s look at the performance-related complexities that teams commonly face with write-heavy workloads and discuss your options for tackling them Let’s look at the performance-related complexities that teams commonly face with write-heavy workloads and discuss your options for tackling them Write-heavy database workloads bring a distinctly different set of challenges than read-heavy ones. For example: Scaling writes can be costly, especially if you pay per operation and writes are 5X more costly than readsLocking can add delays and reduce throughputI/O bottlenecks can lead to write amplification and complicate crash recoveryDatabase backpressure can throttle the incoming load Scaling writes can be costly, especially if you pay per operation and writes are 5X more costly than reads Locking can add delays and reduce throughput I/O bottlenecks can lead to write amplification and complicate crash recovery Database backpressure can throttle the incoming load While cost matters – quite a lot, in many cases – it’s not a topic we want to cover here. Rather, let’s focus on the performance-related complexities that teams commonly face and discuss your options for tackling them. What Do We Mean by “a Real-Time Write Heavy Workload”? First, let’s clarify what we mean by a “real-time write-heavy” workload. We’re talking about workloads that: Ingest a large amount of data (e.g., over 50K OPS)Involve more writes than readsAre bound by strict latency SLAs (e.g., single-digit millisecond P99 latency) Ingest a large amount of data (e.g., over 50K OPS) Involve more writes than reads Are bound by strict latency SLAs (e.g., single-digit millisecond P99 latency) In the wild, they occur across everything from online gaming to real-time stock exchanges. A few specific examples: Internet of Things (IoT) workloads tend to involve small but frequent append-only writes of time series data. Here, the ingestion rate is primarily determined by the number of endpoints collecting data. Think of smart home sensors or industrial monitoring equipment constantly sending data streams to be processed and stored.Logging and Monitoring systems also deal with frequent data ingestion, but they don’t have a fixed ingestion rate. They may not necessarily append only, as well as may be prone to hotspots, such as when one endpoint misbehaves.Online Gaming platforms need to process real-time user interactions, including game state changes, player actions, and messaging. The workload tends to be spiky, with sudden surges in activity. They’re extremely latency sensitive since even small delays can impact the gaming experience.E-commerce and Retail workloads are typically update-heavy and often involve batch processing. These systems must maintain accurate inventory levels, process customer reviews, track order status, and manage shopping cart operations, which usually require reading existing data before making updates.Ad Tech and Real-time Bidding systems require split-second decisions. These systems handle complex bid processing, including impression tracking and auction results, while simultaneously monitoring user interactions such as clicks and conversions. They must also detect fraud in real time and manage sophisticated audience segmentation for targeted advertising.Real-time Stock Exchange systems must support high-frequency trading operations, constant stock price updates, and complex order matching processes – all while maintaining absolute data consistency and minimal latency. Internet of Things (IoT) workloads tend to involve small but frequent append-only writes of time series data. Here, the ingestion rate is primarily determined by the number of endpoints collecting data. Think of smart home sensors or industrial monitoring equipment constantly sending data streams to be processed and stored. Internet of Things (IoT) Logging and Monitoring systems also deal with frequent data ingestion, but they don’t have a fixed ingestion rate. They may not necessarily append only, as well as may be prone to hotspots, such as when one endpoint misbehaves. Logging and Monitoring systems Online Gaming platforms need to process real-time user interactions, including game state changes, player actions, and messaging. The workload tends to be spiky, with sudden surges in activity. They’re extremely latency sensitive since even small delays can impact the gaming experience. Online Gaming E-commerce and Retail workloads are typically update-heavy and often involve batch processing. These systems must maintain accurate inventory levels, process customer reviews, track order status, and manage shopping cart operations, which usually require reading existing data before making updates. E-commerce and Retail Ad Tech and Real-time Bidding systems require split-second decisions. These systems handle complex bid processing, including impression tracking and auction results, while simultaneously monitoring user interactions such as clicks and conversions. They must also detect fraud in real time and manage sophisticated audience segmentation for targeted advertising. Ad Tech and Real-time Bidding Real-time Stock Exchange systems must support high-frequency trading operations, constant stock price updates, and complex order matching processes – all while maintaining absolute data consistency and minimal latency. Real-time Stock Exchange Next, let’s look at key architectural and configuration considerations that impact write performance. Storage Engine Architecture The choice of storage engine architecture fundamentally impacts write performance in databases. Two primary approaches exist: LSM trees and B-Trees. Databases known to handle writes efficiently – such as ScyllaDB, Apache Cassandra, HBase, and Google BigTable – use Log-Structured Merge Trees (LSM). This architecture is ideal for handling large volumes of writes. Since writes are immediately appended to memory, this allows for very fast initial storage. Once the “memtable” in memory fills up, the recent writes are flushed to disk in sorted order. That reduces the need for random I/O. For example, here’s what the ScyllaDB write path looks like: With B-tree structures, each write operation requires locating and modifying a node in the tree – and that involves both sequential and random I/O. As the dataset grows, the tree can require additional nodes and rebalancing, leading to more disk I/O, which can impact performance. B-trees are generally better suited for workloads involving joins and ad-hoc queries. Payload Size Payload size also impacts performance. With small payloads, throughput is good but CPU processing is the primary bottleneck. As the payload size increases, you get lower overall throughput and disk utilization also increases. Ultimately, a small write usually fits in all the buffers and everything can be processed quite quickly. That’s why it’s easy to get high throughput. For larger payloads, you need to allocate larger buffers or multiple buffers. The larger the payloads, the more resources (network and disk) are required to service those payloads. Compression Disk utilization is something to watch closely with a write-heavy workload. Although storage is continuously becoming cheaper, it’s still not free. Compression can help keep things in check – so choose your compression strategy wisely. Faster compression speeds are important for write-heavy workloads, but also consider your available CPU and memory resources. Be sure to look at the compression chunk size parameter. Compression basically splits your data into smaller blocks (or chunks) and then compresses each block separately. When tuning this setting, realize that larger chunks are better for reads while smaller ones are better for writes, and take your payload size into consideration. compression chunk size parameter Compaction For LSM-based databases, the compaction strategy you select also influences write performance. Compaction involves merging multiple SSTables into fewer, more organized files, to optimize read performance, reclaim disk space, reduce data fragmentation, and maintain overall system efficiency. When selecting compaction strategies, you could aim for low read amplification, which makes reads as efficient as possible. Or, you could aim for low write amplification by avoiding compaction from being too aggressive. Or, you could prioritize low space amplification and have compaction purge data as efficiently as possible. For example, ScyllaDB offers several compaction strategies (and Cassandra offers similar ones): several compaction strategies Size-tiered compaction strategy (STCS): Triggered when the system has enough (four by default) similarly sized SSTables.Leveled compaction strategy (LCS): The system uses small, fixed-size (by default 160 MB) SSTables distributed across different levels.Incremental Compaction Strategy (ICS): Shares the same read and write amplification factors as STCS, but it fixes its 2x temporary space amplification issue by breaking huge sstables into SSTable runs, which are comprised of a sorted set of smaller (1 GB by default), non-overlapping SSTables.Time-window compaction strategy (TWCS): Designed for time series data. Size-tiered compaction strategy (STCS): Triggered when the system has enough (four by default) similarly sized SSTables. Size-tiered compaction strategy (STCS): Leveled compaction strategy (LCS): The system uses small, fixed-size (by default 160 MB) SSTables distributed across different levels. Leveled compaction strategy (LCS): Incremental Compaction Strategy (ICS): Shares the same read and write amplification factors as STCS, but it fixes its 2x temporary space amplification issue by breaking huge sstables into SSTable runs, which are comprised of a sorted set of smaller (1 GB by default), non-overlapping SSTables. Incremental Compaction Strategy (ICS): Time-window compaction strategy (TWCS): Designed for time series data. Time-window compaction strategy (TWCS): For write-heavy workloads, we warn users to avoid leveled compaction at all costs. That strategy is designed for read-heavy use cases. Using it can result in a regrettable 40x write amplification. Batching In databases like ScyllaDB and Cassandra, batching can actually be a bit of a trap – especially for write-heavy workloads. If you’re used to relational databases, batching might seem like a good option for handling a high volume of writes. But it can actually slow things down if it’s not done carefully. Mainly, that’s because large or unstructured batches end up creating a lot of coordination and network overhead between nodes. However, that’s really not what you want in a distributed database like ScyllaDB. Here’s how to think about batching when you’re dealing with heavy writes: Batch by the Partition Key: Group your writes by the partition key so the batch goes to a coordinator node that also owns the data. That way, the coordinator doesn’t have to reach out to other nodes for extra data. Instead, it just handles its own, which cuts down on unnecessary network traffic.Keep Batches Small and Targeted: Breaking up large batches into smaller ones by partition keeps things efficient. It avoids overloading the network and lets each node work on only the data it owns. You still get the benefits of batching, but without the overhead that can bog things down.Stick to Unlogged Batches: Considering you follow the earlier points, it’s best to use unlogged batches. Logged batches add extra consistency checks, which can really slow down the write. Batch by the Partition Key: Group your writes by the partition key so the batch goes to a coordinator node that also owns the data. That way, the coordinator doesn’t have to reach out to other nodes for extra data. Instead, it just handles its own, which cuts down on unnecessary network traffic. Batch by the Partition Key: Keep Batches Small and Targeted: Breaking up large batches into smaller ones by partition keeps things efficient. It avoids overloading the network and lets each node work on only the data it owns. You still get the benefits of batching, but without the overhead that can bog things down. Keep Batches Small and Targeted: Stick to Unlogged Batches: Considering you follow the earlier points, it’s best to use unlogged batches. Logged batches add extra consistency checks, which can really slow down the write. Stick to Unlogged Batches: So, if you’re in a write-heavy situation, structure your batches carefully to avoid the delays that big, cross-node batches can introduce. Wrapping Up We offered quite a few warnings, but don’t worry. It was easy to compile a list of lessons learned because so many teams are extremely successful working with real-time write-heavy workloads. Now you know many of their secrets, without having to experience their mistakes. If you want to learn more, here are some firsthand perspectives from teams who tackled quite interesting write-heavy challenges: Zillow: Consuming records from multiple data producers, which resulted in out-of-order writes that could result in incorrect updatesTractian: Preparing for 10X growth in high-frequency data writes from IoT devicesFanatics: Heavy write operations like handling orders, shopping carts, and product updates for this online sports retailer Zillow: Consuming records from multiple data producers, which resulted in out-of-order writes that could result in incorrect updates Zillow: Zillow Tractian: Preparing for 10X growth in high-frequency data writes from IoT devices Tractian Tractian Fanatics: Heavy write operations like handling orders, shopping carts, and product updates for this online sports retailer Fanatics Fanatics Also, take a look at the following video, where we go into even greater depth on these write-heavy challenges and also walk you through what these workloads look like on ScyllaDB.
