22 posts tagged with "tech-blog"

TL;DR​

We optimized two Unicode string helpers — cappedLengthUnicode and cappedByteLengthUnicode — by replacing byte-by-byte utf8proc_char_length calls with a SIMD-based scanning loop. The new implementation processes register-width blocks at a time: pure-ASCII blocks skip in one step, while mixed blocks use bitmask arithmetic to count character starts. Both helpers now share a single parameterized template, eliminating code duplication.

On a comprehensive benchmark matrix covering string lengths from 4 to 1024 bytes and ASCII ratios from 0% to 100%, we measured 2–15× speedups across most configurations, with no regressions on Unicode-heavy inputs. The optimization benefits all callers of these helpers, including the Iceberg truncate transform and various string functions.

Why SIMD, why here​

Velox is a high-performance execution engine, and SIMD is one of the primary ways we deliver that performance across compute-heavy paths. This is not a one-off micro-optimization; it is a principle we apply consistently when the data and algorithms make SIMD a natural fit. String processing is one of those places: it is ubiquitous in query execution and often dominated by byte-wise work that maps well to vectorized scanning.

In this post, we describe how we applied this principle to two Unicode helpers that sit on many string processing paths, and how we balanced speed, correctness, and maintainability.

The old implementation calls utf8proc_char_length on every UTF character. That is robust, but it is inefficient for long ASCII segments where every character has only 1 byte.

What these helpers do​

• cappedLengthUnicode returns the number of UTF-8 characters in a string, stopping early once maxChars is reached.
• cappedByteLengthUnicode returns the byte position of the Nth character, also capped by maxChars. This is the key primitive for truncating strings by character count without re-parsing the entire string.

Consider a UTF-8 string that mixes multi-byte characters and ASCII:

input = "你好abc世界"  // 15 bytes, 7 characters

With maxChars = 3:

• cappedLengthUnicode returns 3.
• cappedByteLengthUnicode returns 7 (the byte position after the third character, "好").

Both need to handle mixed ASCII and multi-byte UTF-8 inputs, invalid bytes, and early-exit behavior. These helpers are called on the Unicode code path — the path used when the input is not known to be pure ASCII at the call site. They are hot in several places, most notably the Iceberg truncate transform, which must find the byte position of the Nth character to truncate partition keys. When we profiled the Iceberg write benchmark, cappedByteLengthUnicode stood out as a clear hot spot.

Byte-by-byte UTF-8 scanning becomes a bottleneck on large, ASCII-heavy strings, but we cannot assume pure ASCII because inputs may mix ASCII and multi-byte characters.

The SIMD design​

We introduced a shared helper, detail::cappedUnicodeImpl, and parameterized it by whether it should return characters or bytes. The key idea is to process the string in SIMD-width blocks using xsimd:

1. ASCII fast path. Load a register-width block and check whether any byte has its high bit set. If not, the whole block is ASCII. We can skip the entire block and advance the counters in one step.
2. Mixed UTF-8 blocks. For a block containing non-ASCII bytes, we count character starts using (byte & 0xC0) != 0x80, convert the predicate into a bitmask, and use popcount to get the number of UTF-8 characters in the block.
3. Finding the Nth character. For cappedByteLengthUnicode, if the maxChars target falls inside a mixed block, we use ctz on the bitmask to locate the Nth character start and add a small inline helper (getUtf8CharLength) to compute the byte length of that UTF-8 character.
4. Remainder handling. For the tail, we keep two loops:
   • A non-ASCII remainder loop that avoids ASCII fast-path mispredicts.
   • A branching ASCII loop that is cheaper than the branchless version for short strings.

The journey: what we tried and learned​

The final implementation emerged through several rounds of iteration and experimentation. Here are the key insights from that process.

Optimize at the core, not the call site​

The initial version added the SIMD fast path directly in the Iceberg truncate function. But the real bottleneck was in cappedByteLengthUnicode itself. Moving the optimization into the shared helper meant every call site benefits — not just Iceberg.

Avoid switching between code paths​

The first SIMD version only accelerated the leading ASCII prefix: it scanned SIMD blocks until it hit the first non-ASCII byte, then fell back to scalar for the rest of the string. This is useless if the first character is non-ASCII but the rest are ASCII. The obvious fix — re-entering the SIMD path after each non-ASCII block — caused a ~20% regression on mixed and Unicode-heavy inputs due to the overhead of switching between SIMD and scalar modes. The solution was a single SIMD loop that handles both ASCII and mixed blocks, avoiding frequent code-path switches and extra branching.

Unify the two helpers​

cappedLengthUnicode and cappedByteLengthUnicode have nearly identical logic — the only difference is what they return. Factoring them into a shared detail::cappedUnicodeImpl<returnByteLength> template eliminated the code duplication and ensures both helpers stay in sync.

Short strings need special care​

After the SIMD path was generalized, the expanded benchmark matrix revealed regressions on very short strings (4 bytes). The fix was simple: skip the SIMD loop entirely when the string is shorter than one register width. Short strings go straight to the scalar path, which avoids the overhead of loading and testing a partial SIMD register.

A branchless getUtf8CharLength​

The scalar remainder loop originally called utf8proc_char_length, which involves several branches. A branchless alternative using std::countl_zero compiles to a handful of instructions:

auto n = std::countl_zero((~static_cast<uint32_t>(c) << 24) | 1);
return (n == 0 || n > 4) ? 1 : n;

This improved performance on the short-string and remainder paths.

Confidence through comprehensive testing​

Since cappedLengthUnicode is widely used across Velox (unlike cappedByteLengthUnicode which is limited to Iceberg and Spark), the change needed high confidence. The PR added tests for multi-byte UTF-8 sequences that straddle SIMD boundaries, continuation bytes in the scalar tail, and strings that are exactly one register width.

Benchmark results​

We added StringCoreBenchmark.cpp to measure both helpers across a matrix of string lengths (4, 16, 64, 256, 1024 bytes) and ASCII ratios (0%, 1%, 25%, 50%, 75%, 99%, 100%). Each benchmark processes 10,000 strings with maxChars = 128.

ASCII ratio indicates the percentage of characters in each string that are single-byte ASCII; the remaining characters are multi-byte UTF-8. The following table highlights representative speedups from the full matrix (measured on an Apple M3):

The small regressions on 4-byte strings are within noise for such tiny inputs. For strings ≥ 16 bytes — which dominate real workloads — speedups range from 2× to over 15×, with the largest gains on mixed-content strings where the SIMD loop handles both ASCII and non-ASCII blocks without falling back to scalar.

We also validated the results on an x86_64 server (Intel Xeon Skylake with AVX-512) and observed similar improvement patterns. A full set of raw numbers is available in the perf data gist and summarized in the PR.

Takeaways​

• Measure first, optimize second. The optimization started from a real hot spot found in the Iceberg write benchmark — not from speculation.
• Optimize the core, not the symptom. Rather than patching the call site, we pushed the improvement into the shared helper so that all callers benefit.
• SIMD is a first-class tool in Velox for performance-critical paths. String processing — ubiquitous in query execution and dominated by byte-wise work — is a natural fit for vectorized scanning.
• Avoid switching between code paths. Re-entering the SIMD path after every non-ASCII block introduced extra branching and switching overhead; a unified loop that handles both ASCII and mixed blocks was consistently faster.
• Branchless is not always faster. A single branchless remainder loop was slower than two tailored loops on short strings.
• Avoid duplication. The two helpers share a single parameterized implementation, keeping behavior consistent and reducing maintenance burden.
• Data-driven iteration. We tried multiple strategies and compared them against a realistic benchmark matrix. We kept the version that improved the common case without regressing Unicode-heavy workloads.
• Comprehensive benchmarks across different string lengths and ASCII mixing ratios are essential for validating correctness-sensitive optimizations.

Acknowledgements​

Special thanks to Masha Basmanova and Yuhta for their guidance and detailed code reviews that shaped the SIMD design and pushed for comprehensive benchmarking.

SQL functions sometimes use regular expressions under the hood in ways that surprise users. Two common examples are the LIKE operator and Spark's split function.

In Presto, split takes a literal string delimiter and regexp_split is a separate function for regex-based splitting. Spark's split, however, always treats the delimiter as a regular expression.

Both LIKE and Spark's split can silently produce wrong results and waste CPU when used with column values instead of constants. Understanding why this happens helps write faster, more correct queries — and helps engine developers make better design choices.

LIKE is not contains​

A very common query pattern is to check whether one string contains another:

SELECT * FROM t WHERE name LIKE '%' || search_term || '%'

This looks intuitive: wrap search_term in % wildcards and you get a "contains" check. But LIKE is not the same as substring matching. LIKE treats _ as a single-character wildcard and % as a multi-character wildcard. If search_term comes from a column and contains these characters, the results are silently wrong:

SELECT url,
       url LIKE '%' || search_term || '%' AS like_result,
       strpos(url, search_term) > 0 AS contains_result
FROM (VALUES
    ('https://site.com/home'),
    ('https://site.com/user_profile'),
    ('https://site.com/username')
) AS t(url)
CROSS JOIN (VALUES ('user_')) AS s(search_term);

              url              | like_result | contains_result
-------------------------------+-------------+----------------
https://site.com/home          |    false    |     false
https://site.com/user_profile  |    true     |     true
https://site.com/username      |    true     |     false

LIKE '%user_%' matches 'https://site.com/username' because _ is a wildcard that matches any single character — in this case, n. But strpos(url, 'user_') > 0 treats _ as a literal underscore and correctly reports that 'https://site.com/username' does not contain the substring 'user_'.

When the pattern is a constant, this distinction is visible and intentional. But when users write x LIKE '%' || y || '%' where y is a column, the values of y may contain _ or % characters — and they will be silently interpreted as wildcards, producing wrong results.

Spark's split treats delimiters as regex​

In Presto, the split function takes a literal string delimiter, while regexp_split is a separate function for regex-based splitting. This distinction makes the intent clear.

Spark's split function, however, always treats the delimiter as a regular expression. Users rarely realize this, and a common pattern is to split a string using a value from another column:

select split(dir_path, location_path)[1] as partition_name from t

Here, a table stores Hive partition metadata: dir_path is the full partition path (e.g., /data/warehouse/db.name/table/ds=2024-01-01) and location_path is the table path (e.g., /data/warehouse/db.name/table). The user wants to strip the table path prefix to get the partition name.

This works for simple paths. But location_path is interpreted as a regular expression, not a literal string. If it contains . — as in db.name — the . matches any character, not a literal dot. Characters like (, ), [, +, *, ?, and $ would also cause wrong results or errors.

A correct alternative that also executes faster uses simple string operations:

IF(starts_with(dir_path, location_path),
   substr(dir_path, length(location_path) + 2)) as partition_name

This is a bit more verbose than split(dir_path, location_path)[1], but it is correct for all inputs and avoids regex compilation entirely.

Performance trap​

Beyond correctness, there is a performance problem. Both LIKE and Spark's split use RE2 as the regex engine. RE2 is fast and safe, but compiling a regular expression can take up to 200x more CPU time than evaluating it.

When the pattern or delimiter is a constant, the regex is compiled once and reused for every row. The cost is negligible. But when the pattern comes from a column, a new regex may need to be compiled for every distinct value. A table with thousands of distinct location_path values means thousands of regex compilations — each one expensive and none of them necessary.

Velox limits the number of compiled regular expressions per function instance per thread of execution via the expression.max_compiled_regexes configuration property (default: 100). When this limit is reached, the query fails with an error.

Tempting but wrong fix​

When users hit this limit, the natural reaction is to ask the engine developers to raise or eliminate the cap. A recent pull request proposed replacing the fixed-size cache with an evicting cache: when the limit is reached, the oldest compiled regex is evicted to make room for the new one.

This sounds reasonable, and the motivation is understandable — users migrating from Spark don't want to rewrite working queries. But it makes things worse:

• It hides the correctness bug. The query no longer fails, so users never discover that their LIKE pattern or split delimiter is being interpreted as a regex and producing wrong results for inputs with special characters.
• It makes the performance problem worse. With thousands of distinct patterns, the cache churns constantly — evicting one compiled regex only to compile another. The query runs, but dramatically slower than necessary, and the user has no indication why. In shared multi-tenant clusters, a single slow query like this can consume excessive CPU and affect other users' workloads.

The error is a feature, not a bug. It is an early warning that catches misuse before it leads to silently wrong results in production and prevents a single query from wasting shared cluster resources.

Right fix​

For users: replace LIKE with literal string operations when checking for substrings. Use strpos(x, y) > 0 or contains(x, y) instead of x LIKE '%' || y || '%'. For Spark's split with literal delimiters, use substr or other string functions that don't involve regex.

For engine developers: optimize the functions to avoid regex when it isn't needed. Velox's LIKE implementation already does this. As described in James Xu's earlier blog post, the engine analyzes each pattern and uses fast paths — prefix match, suffix match, substring search — whenever the pattern contains only regular characters and _ wildcards. For simple patterns, this gives up to 750x speedup over regex. Regex is compiled only for patterns that truly require it, and these optimized patterns are not counted toward the compiled regex limit.

The same approach should be applied to Spark's split function. The engine can check whether the delimiter contains any regex metacharacters. If it doesn't, a simple string search can be used instead of compiling a regex. This would make queries like split(dir_path, location_path) both fast and correct — without users needing to change anything and without removing the safety net for cases that genuinely require regex.

Takeaways​

• LIKE is not contains. The _ and % wildcards can silently corrupt results when the pattern comes from a column.
• Spark's split treats delimiters as regex. Characters like . in column values are interpreted as regex metacharacters, not literal characters. Presto avoids this by separating split (literal) and regexp_split (regex).
• When a query hits the compiled regex limit, the right response is to fix the query, not to raise the limit.
• Engine developers should optimize functions to avoid regex when the input is a plain string, rather than making it easier to misuse regex at scale.

Context​

Strings are ubiquitously used in large-scale analytic query processing. From storing identifiers, names, labels, or structured data (like json/xml), to simply descriptive text, like a product description or the contents of this very blog post, there is hardly a SQL query that does not require the manipulation of string data.

This post describes in more detail how Velox handles columns of strings, the low-level C++ APIs involved and some recent changes made to them, and presents best practices for string usage throughout Velox's codebase.

StringView​

To efficiently enable string operations, Velox provides a specialized physical C++ type, called velox::StringView. The main purpose of this layout, also called a German String Layout, is to optimize common string operations over small strings, or operations that can be performed by only accessing a string's prefix.

StringViews are 16-byte objects that provide a non-owning string-like API that refers to data stored elsewhere in a columnar Velox Vector buffer. The lifetime of a velox::StringView is always tied to the lifetime of the underlying Velox Vector buffer, in the sense that it only provides a pointer (view) into the external buffer. It provides a similar abstraction to std::string_view, with the following additional optimizations:

• Small String Optimization (SSO): strings up to 12 characters are always stored inline, within the object itself. With the assumption that strings are in many cases small, it removes the need for dereferencing the larger data buffer, hopefully reducing cache misses and memory bus traffic.

• Prefix optimization: a 4 character prefix is always stored inline within the StringView object to enable failed comparison to be short circuited, also skipping accesses to the external buffer.

To enable better interoperability with other engines, a few years ago we collaborated with the Arrow community in what resulted to be the BinaryView Arrow format for string data. Ever since, we have seen increased adoption of this format in a series of other systems and libraries.

The Old Unsafe API​

For convenience, in Velox we used to allow developers to implicitly convert velox::StringView into std::string, since many APIs are built based on std::string. For example, one could naively do:

// Given this signature.
void myFunction(const std::string& input) {
  // ...
}

// All would silently result in a string copy
// (valueAt() returns a velox::StringView):

myFunction(stringVector->valueAt(0));

std::string myStr = stringVector->valueAt(0);

std::unordered_map<std::string, size_t> myMap;
myMap[stringVector->valueAt(0)] = 1;

While these usages seem harmless, they would all result in a full string copy, and potentially a memory allocation, depending on the string size. We have found this anti-pattern to be commonly used; in most cases, the developer did not intend for a copy to be performed, and this behavior silently added unnecessary overhead.

What Changed​

To prevent such inadvertent string copies, in #15946 we removed implicit conversion from velox::StringView to std::string. You can still do so if needed, but you now have to explicitly state it, for clarity:

// Compilation failure from now on:
myFunction(stringVector->valueAt(0));

// Ok:
myFunction(std::string(stringVector->valueAt(0)));

Instead, we are now making conversion from velox::StringView to std::string_view available implicitly. std::string_view's are non-owning, so constructing them based on a pointer and size is essentially free.

A series of PR were landed before this change to clean up any dependencies on the old behavior throughout Velox's codebase, making them explicitly defined when needed.

folly::StringPiece​

folly::StringPiece's are now superseded by std::string_view. As part of this work, we have also cleaned up and removed every usage of folly::StringPiece in Velox; do not use folly::StringPiece in future code in Velox. Use std::string_view or velox::StringView instead.

If you need to use a library that takes a folly::StringPiece, for example, folly::parseJson(), you can pass a std::string_view instead, since the std::string_view => folly::StringPiece conversion can be done implicitly.

The RValue Gotcha​

A slightly unintuitive gotcha of velox::StringView is the handling of rvalues and temporary values. Since velox::StringView may contain inlined strings, if you take a pointer to the string contents, the string happened to be small and inlined, and the velox::StringView object is destroyed, you would end up with a dangling pointer. For example:

// Unsafe: would result in a std::string_view pointing to
// the temporary object that gets destroyed right away.
std::string_view sv = stringVector->valueAt(0);
LOG(INFO) << "Happily would have crashed: " << sv;

To prevent these unsafe usages, the rvalue-based version of these conversions, both explicit and implicit, and for both std::string and std::string_view, are disabled. This means that the snippets above will give you a compilation error.

The safe behavior would be to create a non-temporary object in the stack:

velox::StringView sv1 = stringVector->valueAt(0);
std::string_view sv2 = sv1; // all good, as long as sv2 is not used after
                             // sv1 is destroyed.

Best Practices​

To summarize, from now on, the conversions:

• velox::StringView => std::string: won't be done implicitly anymore. This now needs to be explicitly stated, for clarity.
• velox::StringView => std::string_view: can be done implicitly.
• velox::StringView => folly::StringPiece: do not use. If needed for legacy external libraries, pass a std::string_view instead.
• rvalue velox::StringView&& => any other type: compilation error to minimize risk of dangling pointer.

Please reach out on Slack if you have any questions.

TL;DR​

Velox Task Barriers provide a synchronization mechanism that not only enables efficient task reuse, important for workloads such as AI training data loading, but also delivers the strict sequencing and checkpointing semantics required for streaming workloads.

By injecting a barrier split, users guarantee that no subsequent data is processed until the entire DAG is flushed and the synchronization signal is unblocked. This capability serves two critical patterns:

1. Task Reuse: Eliminates the overhead of repeated task initialization and teardown by safely reconfiguring warm tasks for new queries. This is a recurring pattern in AI training data loading workloads.

2. Streaming Processing: Enables continuous data handling with consistent checkpoints, allowing stateful operators to maintain context across batches without service interruption.

See the Task Barrier Developer Guide for implementation details.

The Problem​

Creating a new Velox task for every data request introduces overhead:

• Memory pool initialization and configuration
• Pipeline and operator construction
• Connector and data source setup
• Thread pool registration

For high-throughput workloads that process thousands of requests, this overhead may accumulate and impact overall system performance.

The Solution: Task Barrier​

Task Barrier introduces a lightweight synchronization mechanism that:

1. Drains all in-flight data - Ensures all buffered data flows through the pipeline to completion.

2. Resets stateful operators - Clears internal state in operators like aggregations, joins, and sorts.

3. Preserves task infrastructure - Keeps memory pools, thread registrations, and connector sessions intact.

How It Works​

The barrier is triggered by a special BarrierSplit injected into the task's source operators via requestBarrier(). When received:

1. Source operators propagate the barrier signal downstream.
2. Each operator calls finishDriverBarrier() to flush and reset its state.
3. Once all drivers complete barrier processing, the task is ready for a new split.

Workflow Example (Pseudo-code)​

The following example demonstrates the usage pattern: the user adds splits, requests a barrier, and then must continue to drain results until the barrier is reached. This ensures all in-flight data leaves the pipeline, allowing the barrier to complete.

// 1. Initialize the task infrastructure once
auto task = Task::create(config, memoryPool);

// 2. Loop to process multiple batches using the same task
while (scheduler.hasMoreRequests()) {

  // A. Get new data splits (e.g., from a file or query)
  auto splits = scheduler.getNextSplits();

  // B. Add splits to the running task
  task->addSplits(sourcePlanNodeId, splits);

  // C. Request a barrier to mark the end of this batch.
  // This injects a BarrierSplit that flows behind the data.
  // Note: Any splits added to the task AFTER this call will be blocked
  // and will not be processed until this barrier is fully cleared.
  auto barrierFuture = task->requestBarrier();

  // D. Drain the pipeline until the barrier is passed.
  // Critical: We must keep pulling results so the task can flush
  // its buffers and reach the barrier.
  while (!barrierFuture.isReady()) {
    auto result = task->next(); // Get next result batch
    if (result != nullptr) {
      userOutput.consume(result);
    }
  }

  // E. Barrier reached.
  // The task has been reset (operators cleared, buffers flushed).
  // It is now ready for the next iteration immediately.
}

Operator Support​

Task Barrier requires operators to implement finishDriverBarrier() to properly reset their state:

• TableScan: Clears current split and reader state.
• HashBuild/HashProbe: Resets hash tables and probe state.
• Aggregation: Flushes partial aggregates and clears accumulators.
• OrderBy: Outputs sorted data and resets sort buffer.
• Exchange: Drains exchange queues.

Performance Impact​

By eliminating repeated task creation overhead, Task Barrier provides significant performance improvements for workloads with:

• Short-lived Queries: Specifically those that recur frequently with the exact same plan DAG, avoiding the need for repeated reconstruction.
• AI/ML Pipelines: High-frequency data loading requests where task initialization would otherwise dominate execution time.
• Iterative Processing: Queries that can benefit from "warm" caches and pre-allocated memory pools.

Learn More​

For detailed implementation information, API reference, and operator-specific behavior, see the Task Barrier documentation.

References​

• Data Ingestion for Machine Learning Training at Meta
• Flink Stateful Streaming Processing
• Velox Task Model
• Velox Task Barrier

TL;DR​

Velox is a fully vectorized execution engine[1]. Its internal columnar memory layout enhances cache locality, exposes more inter-instruction parallelism to CPUs, and enables the use of SIMD instructions, significantly accelerating large-scale query processing.

However, some operators in Velox utilize a hybrid layout, where datasets can be temporarily converted to a row-oriented format. The OrderBy operator is one example, where our implementation first materializes the input vectors into rows, containing both sort keys and payload columns, sorts them, and converts the rows back to vectors.

In this article, we explain the rationale behind this design decision and provide experimental evidence for its implementation. We show a prototype of a hybrid sorting strategy that materializes only the sort-key columns, reducing the overhead of materializing payload columns. Contrary to expectations, the end-to-end performance did not improve—in fact, it was even up to 3× slower. We present the two variants and discuss why one is counter-intuitively faster than the other.

Row-based vs. Non-Materialized​

Row-based Sort​

The OrderBy operator in Velox’s current implementation uses a utility called SortBuffer to perform the sorting, which consists of three stages:

1. Input Stage: Serializes input Columnar Vectors into a row format, stored in a RowContainer.
2. Sort Stage: Sorts based on keys within the RowContainer.
3. Output Stage: Extract output vectors column by column from the RowContainer in sorted order.

While row-based sorting is more efficient than column-based sorting[2,3], what if we only materialize the sort key columns? We could then use the resulting sort indices to gather the payload data into the output vectors directly. This would save the cost of converting the payload columns to rows and back again. More importantly, it would allow us to spill the original vectors directly to disk rather than first converting rows back into vectors for spilling.

Non-Materialized Sort​

We have implemented a non-materializing sort strategy designed to improve sorting performance. The approach materializes only the sort key columns and their original vector indices, which are then used to gather the corresponding rows from the original input vectors into the output vector after the sort completes. It changes the SortBuffer to NonMaterizedSortBuffer, which consists of three stages:

1. Input Stage: Holds the input vector (its shared pointer) in a list, serializes key columns and additional index columns (VectorIndex and RowIndex) into rows, stored in a RowContainer.
2. Sort Stage: Sorts based on keys within the RowContainer.
3. Output Stage: Extracts the VectorIndex and RowIndex columns, uses them together to gather the corresponding rows from the original input vectors into the output vector.

In theory, this should have significantly reduced the overhead of materializing payload columns, especially for wide tables, since only sorting keys are materialized. However, the benchmark results were the exact opposite of our expectations. Despite successfully eliminating expensive serialization overhead and reducing the total instruction the end-to-end performance was 3x times slower.

Benchmark Result​

To validate the effectiveness of our new strategy, we designed a benchmark with a varying number of payload columns:

• Inputs: 1000 Input Vectors, 4096 rows per vector.
• Number of payload columns: 64, 128, 256.
• L2 cache: 80 MiB, L3 cache: 108 MiB.

The benchmark results confirm that Row-based Sort is the superior strategy, delivering a 1.9x to 3.9x overall speedup compared to Columnar Sort. While Row-based Sort incurs a significantly higher upfront cost during the Input phase (peaking at 104s), it maintains a highly stable and efficient Output phase (maximum 32s). In contrast, Columnar Sort suffers from severe performance degradation in the Output phase as the payload increases, with execution times surging from 42s to 283s, resulting in a much slower total execution time despite its negligible input overhead.

To identify the root cause of the performance divergence, we utilized perf stat to analyze micro-architectural efficiency and perf mem to profile memory access patterns during the critical Output phase.

The results reveal a stark contrast in CPU utilization. Although the Row-based approach executes 17% more instructions (due to serialization overhead), it maintains a high IPC of 2.4, indicating a fully utilized pipeline. In contrast, the Columnar approach suffers from a low IPC of 0.82, meaning the CPU is stalled for the majority of cycles. This is directly driven by the 35x difference in LLC Load Misses, which forces the Columnar implementation to fetch data from main memory repeatedly. The memory profile further confirms this bottleneck: Columnar mode is severely latency-bound, spending 38.1% of its execution time waiting for DRAM (RAM Hit) and experiencing significant congestion in the Line Fill Buffer (18.9% LFB Hit), while Row-based mode effectively utilizes the cache hierarchy.

The Memory Access Pattern​

Why does the non-materializing sort, specifically its gather method, cause so many cache misses? The answer lies in its memory access pattern. Since Velox is a columnar engine, the output is constructed column by column. For each column in an output vector, the gather process does the following:

1. It iterates through all rows of the current output vector.
2. For each row, locate the corresponding input vector via the sorted vector index.
3. Locates the source row in the corresponding input vector.
4. Copies the data from that single source cell to the target cell.

The sorted indices, by nature, offer low predictability. This forces the gather operation for a single output column to jump unpredictably across as many as different input vectors, fetching just one value from each. This random access pattern has two devastating consequences for performance.

First, at the micro-level, every single data read becomes a "long-distance" memory jump. The CPU's hardware prefetcher is rendered completely ineffective by this chaotic access pattern, resulting in almost every lookup yielding a cache miss.

Second, at the macro-level, the problem compounds with each column processed. The sheer volume of data touched potentially exceeds the size of the L3 cache. This ensures that by the time we start processing the next payload column, the necessary vectors have already been evicted from the cache. Consequently, the gather process must re-fetch the same vector metadata and data from main memory over and over again for each of the 256 payload columns. This results in 256 passes of cache-thrashing, random memory access, leading to a catastrophic number of cache misses and explaining the severe performance degradation.

In contrast, Velox’s current row-based approach serializes all input vectors into rows, with each allocation producing a contiguous buffer that holds a subset of those rows. Despite the serialization, the row layout preserves strong locality when materializing output vectors: once rows are in the cache, they can be used to extract multiple output columns. This leads to much better cache-line utilization and fewer cache misses than a columnar layout, where each fetched line often yields only a single value per column. Moreover, the largely sequential scans over contiguous buffers let the hardware prefetcher operate effectively, boosting throughput even in the presence of serialization overhead.

Conclusion​

This study reinforces the core principle of performance engineering: Hardware Sympathy. Without understanding the characteristics of the memory hierarchy and optimizing for it, simply reducing the instruction count usually does not guarantee better performance.

Reference​

• [1] https://velox-lib.io/blog/velox-primer-part-1/
• [2] https://duckdb.org/pdf/ICDE2023-kuiper-muehleisen-sorting.pdf
• [3] https://duckdb.org/2021/08/27/external-sorting

Background​

Efficiently merging sorted data partitions at scale is crucial for a variety of training data preparation workloads, especially for Generative Recommenders (GRs) a new paradigm introduced in the paper Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations. A key requirement is to merge training data across partitions—for example, merging hourly partitions into daily ones—while ensuring that all rows sharing the same primary key are stored consecutively. Training data is typically partitioned and bucketed by primary key, with rows sharing the same key stored consecutively, so merging across partitions essentially becomes a multi-way merge problem.

Normally, Apache Spark can be used for this sort-merge requirement — for example, via CLUSTER BY. However, training datasets for a single job can often reach the PB scale, which in turn generates shuffle data at PB scale. Although we typically apply bucketing and ordering by key when preparing training data in production, Spark can eliminate the shuffle when merging training data from multiple hourly partitions. However, each Spark task can only read the files planned from various partitions within a split sequentially, placing them into the sorter and spilling as needed. Only after all files have been read does Spark perform a sort-merge of the spilled files. This process produces a large number of small spill files, which further degrades efficiency.

Moreover, Spark’s spill is row-based with a low compression ratio, resulting in approximately 4 times amplification compared to the original columnar training data in the data lake. These factors significantly degrade task stability and performance. Velox has a LocalMerge operator that can be introduced into Apache Spark via Gluten or PySpark on Velox.

Note: To keep the focus on merging, the remainder of this article also assumes that each partition’s training data is already sorted by primary key—a common setup in training data pipelines.

LocalMerge Operator​

The LocalMerge operator consolidates its sources’ outputs into a single, sorted stream of rows. It runs single-threaded, while its upstream sources may run multi-threaded within the same task, producing multiple sorted inputs concurrently. For example, when merging 24 hourly partitions into a single daily partition (as shown in the figure below), the merge plan fragment is split into two pipelines:

• Pipeline 0: contains two operators, TableScan and CallbackSink. 24 drivers are instantiated to scan the 24 hourly partitions.
• Pipeline 1: contains only a single operator, LocalMerge, with one driver responsible for performing the sort merge.

A CallbackSink operator is installed at the end of each driver in Pipeline 0. It pushes the TableScan operator’s output vectors into the queues backing the merge streams. Inside LocalMerge, a TreeOfLosers performs a k-way merge over the 24 merge streams supplied by the Pipeline 0 drivers, producing a single, globally sorted output stream.

Multi-Round Spill Merge​

Although LocalMerge minimizes comparisons during merging, preserves row-ordering guarantees, and cleanly isolates the single-threaded merge from the multi-threaded scan phase for predictable performance, it can cause substantial memory pressure—particularly in training-data pipelines. In these workloads, extremely wide tables are common, and even after column pruning, thousands of columns may remain.

Moreover, training data is typically stored in PAX-style formats such as Parquet, ORC, or DRWF. Using Parquet as an example, the reader often needs to keep at least one page per column in memory. As a result, simply opening a Parquet file with thousands of columns can consume significant memory even before any merging occurs. Wide schemas further amplify per-column metadata, dictionary pages, and decompression buffers, inflating the overall footprint. In addition, the k-way merge must hold input vectors from multiple sources concurrently, which drives peak memory usage even higher.

To cap memory usage and avoid OOM when merging a large number of partitions, we extend LocalMerge to process fewer local sources at a time, leverage existing spill facilities to persist intermediate results, and introduce lazy-start activation for merge inputs. Using the case of merging 24 hourly partitions into a single daily partition, the process is organized into two phases:

Phase 1

1. Break the scan-and-merge into multiple rounds (e.g., 3 rounds).
2. In each round, lazily start a limited number of drivers (e.g., drivers 0–7, eight at a time).
3. The started drivers scan data and push it into the queues backing their respective merge streams.
4. Perform an in-memory k-way merge and spill the results, producing a spill-file group (one or more spill files per group).
5. After all inputs from drivers 0–7 are consumed and spilled, the drivers will be closed, and close the file streams opened by their TableScan operators, and release associated memory.
6. Repeat the above steps for the remaining rounds (drivers 8–15, then drivers 16–23), ensuring peak memory stays within budget.

Phase 2

1. Create a concatenated file stream for each spill-file group produced in Phase 1.
2. Schedule one async callback for each concatenated stream to prefetch and push data into a merge stream.
3. Merge the outputs of the three merge streams using a k-way merge (e.g., a loser-tree), and begin streaming the final, globally sorted results to downstream operators.
4. The output batch rows is limited adaptively by estimating row size from the merge streams which use the averaged row size from the first batch.

How To Use​

Set local_merge_spill_enabled to true to enable spilling for the LocalMerge operator (it is false by default). Then, set local_merge_max_num_merge_sources to control the number of merge sources per round according to your memory management strategy.

Note: An executor must be configured for spilling, as it would schedule an asynchronous callback for each concatenated stream to prefetch data and push it into the merge stream.

Future Work​

The number of merge sources is adjusted dynamically based on available memory, rather than being determined by the local_merge_max_num_merge_sources parameter. The process starts with a small number of sources, such as 2, and incrementally increases this number for subsequent rounds (e.g., to 4) as long as sufficient memory is available. The number of sources stops increasing once it reaches a memory-constrained limit.

Acknowledgements​

Thanks to Xiaoxuan Meng and Pedro Pederia for their guidance, review, and brainstorming. I also appreciate the excellent collaboration and work from my colleagues, Xiang Yao, Gang Wang, and Weixin Xu.

In this post, I’ll share how we unblocked shared library builds in Velox, the challenges we encountered with our large CMake build system, and the creative solution that let us move forward without disrupting contributors or downstream users.

The State of the Velox Build System​

Velox’s codebase was started in Meta’s internal monorepo, which still serves as a source of truth. Changes from pull requests in the Github repository are not merged directly via the web UI. Instead, the changes are imported into the internal review and CI tool Phabricator, as a ‘diff’. There, it has to pass an additional set of CI checks before being merged into the monorepo and in turn, exported to Github as a commit on the ‘main’ branch.

Internally, Meta uses the Buck2 build system which, like its predecessor Buck, promotes small, granular build targets to improve caching and distributed compilation. Externally, however, the open-source community relies on CMake as a build system. When the CMake build system was first added, its structure mirrored the granular nature of the internal build targets.

The result was hundreds of targets: over one hundred library targets, built as static libraries, and more than two hundred executables for tests and examples. While Buck(2) is optimized for managing such granular targets, the same can not be said for CMake. Each subdirectory maintained its own CMakeLists.txt, and header includes were managed through global include_directories(), with no direct link between targets and their associated header files. This approach encouraged tight coupling across module boundaries. Over the years, dependencies accumulated organically, resulting in a highly interconnected, and in parts cyclic, dependency graph.

Static Linking vs. Shared Linking​

The combination of 300+ targets and static linking resulted in a massive build tree, dozens of GiB when building in release mode and several times larger when built with debug information. This grew to the point where we had to provision larger CI runners just to accommodate the build tree in debug mode!

Individual test executables could reach several GiB, making it impractical to transfer the executables between runners to parallelize the execution of the tests across different CI runners. Significantly delaying CI feedback for developers when pushing changes to a PR.

This size is the result of each static library containing a full copy of all of its dependencies' object files. With over 100+ strongly connected libraries, we end up with countless redundant copies of the same objects scattered throughout the build tree, blowing up the total size.

CMake usually requires the library dependency graph to be acyclic (a DAG). However it allows circular dependencies between static libraries, because it’s possible to adjust the linker invocation to work around possible issues with missing symbols.

For example, say library A depends on foo() from library B, and B in turn depends on bar() defined in A. If we link them in order A B, the linker will find foo() in B for use in A but will fail to find bar() for use in B. This happens because the linker processes the libraries from left to right and symbols that are not (yet) required are effectively ignored.

The trick is now simply to repeat the entire circular group in the linker arguments A B A B. Previously ignored symbols are now in the list of required symbols and can be resolved when the linker processes the repeated libraries.

However, there is no workaround for shared libraries. So the moment we attempted to build Velox as a set of shared libraries, CMake failed to configure. The textbook solution would involve refactoring dependencies between targets, explicitly marking dependencies - as PUBLIC( meaning forwarded to both direct and transitive dependents) or PRIVATE, ( required only for this target at build time), and manually breaking cycles. But with hundreds of targets and years of organic growth, this became an unmanageable task.

Solution Attempt 1: Automate​

Manually addressing this was unmanageable so, our first instinct as Software Engineers was, of course, to automate the process. I wrote a python package that parses all CMakelists.txt files in the repository using Antlr4, tracks which source files and headers belong to which target and reconstructs the library dependency graph. Based on the graph and whether headers are included from a source or a header file, the tool grouped the linked targets as either PRIVATE or PUBLIC.

While this worked locally and made it easier to resolve the circular dependencies, it proved far too brittle and disruptive to implement at scale. Velox's high development velocity meant many parts of the code base changed daily, making it impossible to land modifications across over 250+ CMake files in a timely manner while also allowing for thorough review and testing. Maintaining this carefully constructed DAG of dependencies would also have required an unsustainable amount of specialized reviewer attention, time that could be spent much more productively elsewhere.

Solution: Monolithic Library via CMake Wrappers​

Given the practical constraints we needed a solution that would unblock shared builds without disrupting the ongoing development. To achieve this we used a common pattern of larger CMake code bases, project specific wrapper functions. A special thanks to Marcus Hanwell for introducing me to the pattern he successfully used in VTK.

We created wrapper functions for a selection of the CMake commands that create and manage build targets: add_library became velox_add_library, target_link_libraries turned into velox_link_libraries etc. These functions just forward the arguments to the wrapped core command, unless the option to build Velox as a monolithic library is set. In the latter case instead of ~100 only a single library target is created.

Each call to velox_add_library just adds the relevant source files to the main velox target. To ensure drop-in compatibility, for every original library target, the wrapper creates an ALIAS target referring to velox. This way, all existing code, downstream consumers, old and new test targets, can continue to link to their accustomed velox_xyz targets, which effectively point to the same monolithic library. There is no need to adjust CMake files throughout the codebase, keeping the migration quick and entirely transparent to developers and users.

Some special-case libraries, such as those involving the CUDA driver, still needed to be built separately for runtime or integration reasons. The wrapper logic allows exemptions for these, letting them be built as standalone libraries even in monolithic mode.

Here's a simplified version of the relevant logic:

function(velox_add_library TARGET)
  if(VELOX_MONO_LIBRARY)
      if(TARGET velox)
        # Target already exists, append sources to it.
        target_sources(velox PRIVATE ${ARGN})
      else()
        add_library(velox ${ARGN})
      endif()
      # create alias for compatibility
      add_library(${TARGET} ALIAS velox)
  else()
    # Just call add_library as normal
    add_library(${TARGET} ${ARGN})
  endif()
endfunction()

This new, monolithic library obviously doesn’t suffer from cyclic dependencies, as it only has external dependencies. So after implementing the wrapper functions, building Velox as a shared library only required some minor adjustments!

We have now switched the default configuration to build the monolithic library and transitioned to build the shared library in our main PR CI and macOS builds.

Result​

The results of our efforts can be clearly seen in our Build Metrics Report with the size of the resulting binaries being the biggest win as seen in Table 1. The workflow that collects these metrics uses our regular CI image with GCC 12.2 and ld 2.38.

The new executables built against the shared libvelox.so are a mere ~4% of their previous size! This unlocks improvements across the board both for the local developer experience, packaging/deployments for downstream users as well as CI.

Table 1: Total size of test, fuzzer and benchmark executables

Less impressive but still notable is the reduction of the debug build time by almost two hours¹. The improvement is entirely explained by the much faster link time for the shared build of ~30 minutes compared to the 2 hours and 20 minutes it takes to link the static library.

Table 2: Total time spent on compilation and linking across all cores

In addition to unblocking shared builds, the wrapper function pattern offers further benefits for future feature additions or build-system wide adjustments with minimal disruption. We already use them to automate installation of header files along the Velox libraries, with no changes to the targets themselves.

The wrapper functions will also help with defining and enforcing the public Velox API by allowing us to manage header visibility, component libraries like GPU and Cloud extensions and versioning in a consistent and centralized manner.

TL;DR​

This post describes the design principles and software components for extending Velox with hardware acceleration libraries like NVIDIA's cuDF. Velox provides a flexible execution model for hardware accelerators, and cuDF's data structures and algorithms align well with core components in Velox.

How Velox and cuDF Fit Together​

Extensibility is a key feature in Velox. The cuDF library integrates with Velox using the DriverAdapter interface to rewrite query plans for GPU execution. The rewriting process allows for operator replacement, operator fusion and operator addition, giving cuDF a lot of freedom when preparing a query plan with GPU operators. Once Velox converts a query plan into executable pipelines, the cuDF DriverAdapter rewrites the plan to use GPU operators:

Generally the cuDF DriverAdapter replaces operators one-to-one. An exception happens when a GPU operator is not yet implemented in cuDF, in which case a CudfConversion operator must be added to transfer from GPU to CPU, fallback to CPU execution, and then transfer back to GPU. For end-to-end GPU execution where cuDF replaces all of the Velox CPU operators, cuDF relies on Velox's pipeline-based execution model to separate stages of execution, partition the work across drivers, and schedule concurrent work on the GPU.

Velox and cuDF benefit from a shared data format, using columnar layout and Arrow-compatible buffers. The CudfVector type in the experimental cuDF backend inherits from Velox's RowVector type, and manages a std::unique_ptr to a cudf::table (link) which owns the GPU-resident data. CudfVector also maintains an rmm::cuda_stream_view (link) to ensure that asynchronously launched GPU operations are stream-ordered. Even when the data resides in GPU memory, CudfVector allows the operator interfaces to follow Velox's standard execution model of producing and consuming RowVector objects.

The first set of GPU operators in the experimental cuDF backend include TableScan, HashJoin, HashAggregation, FilterProject, OrderBy and more. cuDF's C++ API covers many features beyond this initial set, including broad support for list and struct types, high-performance string operations, and configurable null handling. These features are critical for running workloads in Velox with correct semantics for Presto and Spark users. Future work will integrate cuDF support for more Velox operators.

GPU execution: fewer drivers and larger batches​

Compared to the typical settings for CPU execution, GPU execution with cuDF benefits from a lower driver count and larger batch size. Whereas Velox CPU performs best with ~1K rows per batch and driver count equal to the number of physical CPU cores, we have found Velox's cuDF backend to perform best with ~1 GiB batch size and 2-8 drivers for a single GPU. cuDF GPU operators launch one or more device-wide kernels and a single driver may not fully utilize GPU compute. Additional drivers improve throughput by queuing up more GPU work and avoiding stalling during host-to-device copies. Adding more drivers increases memory pressure linearly, and query processing throughput saturates once GPU compute is fully utilized. Please check out our talk, “Accelerating Velox with RAPIDS cuDF” from VeloxCon 2025 to learn more.

Call to Action​

The experimental cuDF backend in Velox is under active development. It is implemented entirely in C++ and does not require CUDA programming knowledge. There are dozens more APIs in cuDF that can be integrated as Velox operators, and many design areas to explore such as spilling, remote storage connectors, and expression handling.

If you're interested in bringing GPU acceleration to the Velox ecosystem, please join us in the #velox-oss-community channel on the Velox Slack workspace. Your contributions will help push the limits of performance in Presto, Spark and other tools powered by Velox.

Background​

Velox depends on several libraries. Some of these dependencies include open-source libraries from Meta, including Folly and Facebook Thrift. These libraries are in active development and also depend on each other, so they all have to be updated to the same version at the same time.

Updating these dependencies typically involves modifying the Velox code to align with any public API or semantic changes in these dependencies. However, a recent upgrade of Folly and Facebook Thrift to version v2025.04.28.00 caused a SEGFAULT only in one unit test in Velox named velox_functions_remote_client_test.

Investigation​

We immediately put on our gdb gloves and looked at the stack traces. This issue was also reproducible in a debug build. The SEGFAULT occurred in Facebook Thrift's ThriftServer Class during it's initialization but the offending call was invoking a destructor of a certain handler. However, the corresponding source code was pointing to an invocation of a different function. And this code was present inside a Facebook Thrift header called AsyncProcessor.h.

This handler (RemoteServer) was implemented in Velox as a Thrift definition. Velox compiled this thrift file using Facebook Thrift, and the generated code was using the ServerInterface class in Facebook Thrift. This ServerInterface class was further extended from both the AsyncProcessorFactory and ServiceHandlerBase interfaces in Facebook Thrift.

One of the culprits resulting in SEGFAULTs in the past was the conflict due to the usage of Apache Thrift and Facebook Thrift. However, this was not the root cause this time because we were able to reproduce this issue by just building the test without the Apache Thrift dependency installed. We were entering a new territory to investigate, and we were not sure where to start.

We then compiled an example called EchoService in Facebook Thrift that was very similar to the RemoteServer, and it worked. Then we copied and compiled the Velox RemoteServer in Facebook Thrift and that worked too! So the culprit was likely in the compilation flags, which likely differed between Facebook Thrift and Velox. We enabled the verbose logging for both builds and were able to spot one difference. We saw the GCC coroutines flag being used in the Facebook Thrift build.

We were also curious about the invocation of the destructor instead of the actual function. We put our gdb gloves back on and dumped the entire vtable for the RemoteServer class and its base classes. The vtables were different when it was built in Velox vs. Facebook Thrift. Specifically, the list of functions inherited from ServiceHandlerBase was different.

The vtable for the RemoteServer handler in the Velox build had the following entries:

folly::SemiFuture<folly::Unit> semifuture_onStartServing()
folly::SemiFuture<folly::Unit> semifuture_onStopRequested()
Thunk ServiceHandlerBase::~ServiceHandlerBase

The vtable for the RemoteServer handler in the Facebook Thrift build had the following entries:

folly::coro::Task<void> co_onStartServing()
folly::coro::Task<void> co_onStopRequested()
folly::SemiFuture<folly::Unit> semifuture_onStartServing()
folly::SemiFuture<folly::Unit> semifuture_onStopRequested()
Thunk ServiceHandlerBase::~ServiceHandlerBase

Tying up both pieces of evidence, we could conclude that Velox generated a different vtable structure compared to what Facebook Thrift (and thus ThriftServer) was built with. Looking around further, we noticed that the ServiceHandlerBase was conditionally adding functions based on the coroutines compile flag that influences the FOLLY_HAS_COROUTINES macro from the portability header.

class ServiceHandlerBase {
  ....
 public:
#if FOLLY_HAS_COROUTINES
  virtual folly::coro::Task<void> co_onStartServing() { co_return; }
  virtual folly::coro::Task<void> co_onStopRequested() {
    throw MethodNotImplemented();
  }
#endif

  virtual folly::SemiFuture<folly::Unit> semifuture_onStartServing() {
#if FOLLY_HAS_COROUTINES
    return co_onStartServing().semi();
#else
    return folly::makeSemiFuture();
#endif
  }

As a result, the ThriftServer would access an incorrect function (~ServiceHandlerBase destructor at offset 3 in the first vtable above) instead of the expected initialization function (semifuture_onStartServing at offset 3 in the second vtable above), thus resulting in a SEGFAULT. We recompiled the Facebook Thrift dependency for Velox with the coroutines compile flag disabled, and the test passed.

At the end of the previous article, we were halfway through running our first distributed query:

SELECT l_partkey, count(*) FROM lineitem GROUP BY l_partkey;

We discussed how a query starts, how tasks are set up, and the interactions between plans, operators, and drivers. We have also presented how the first stage of the query is executed, from table scan to partitioned output - or the producer side of the shuffle.  

In this article, we will discuss the second query stage, or the consumer side of the shuffle.

Shuffle Consumer​

As presented in the previous post, on the first query stage each worker reads the table, then produces a series of information packets (SerializedPages) intended for different workers of stage 2. In our example, the lineitem table has no particular physical partitioning or clustering key. This means that any row of the table can have any value of l_partkey in any of the files forming the table. So in order to group data based on l_partkey, we first need to make sure that rows containing the same values for l_partkey are processed by the same worker – the data shuffle at the end of stage 1.

The overall query structure is as follows:

The query coordinator distributes table scan splits to stage 1 workers in no particular order. The workers process these and, as a side effect, fill destination buffers that will be consumed by stage 2 workers. Assuming there are 100 stage 2 workers, every stage 1 Driver has its own PartitionedOutput which has 100 destinations. When the buffered serializations grow large enough, these are handed off to the worker's OutputBufferManager. 

Now let’s focus on the stage 2 query fragment. Each stage 2 worker has the following plan fragment: PartitionedOutput over Aggregation over LocalExchange over Exchange.  

Each stage 2 Task corresponds to one destination in the OutputBufferManager of each stage 1 worker. The first stage 2 Tasks fetches destination zero from all the stage 1 Tasks. The second stage 2 fetches destination 1 from the first stage Tasks and so on.  Everybody talks to everybody else. The shuffle proceeds without any central coordination.

But let's zoom in to what actually happens at the start of stage 2.

The plan has a LocalExchange node after the Exchange. This becomes two pipelines: Exchange and LocalPartition on one side, and LocalExchange, HashAggregation, and PartitionedOutput on the other side. 

The Velox Task is intended to be multithreaded, with typically 5 to 30 Drivers. There can be hundreds of Tasks per stage, thus amounting to thousands of threads per stage. So, each of the 100 second stage workers is consuming 1/100 of the total output of the first stage. But it is doing this in a multithreaded manner, with many threads consuming from the ExchangeSource. We will explain this later.

In order to execute a multithreaded group by, we can either have a thread-safe hash table or we can partition the data in n disjoint streams, and then proceed aggregating each stream on its own thread. On a CPU, we almost always prefer to have threads working on their own memory, so data will be locally partitioned based on l_partkey using a local exchange. CPUs have complex cache coherence protocols to give observers a consistent ordered view of memory, so reconciliation after many cores have written the same cache line is both mandatory and expensive. Strict memory-to-thread affinity is what makes multicore scalability work. 

LocalPartition and LocalExchange​

To create efficient and independent memory access patterns, the second stage reshuffles the data using a local exchange. In concept, this is like the remote exchange between tasks, but is scoped inside the Task. The producer side (LocalPartition) calculates a hash on the partitioning column l_partkey, and divides this into one destination per Driver in the consumer pipeline. The consumer pipeline has a source operator LocalExchange that reads from the queue filled by LocalPartition. See LocalPartition.h for details. Also see the code in Task.cpp for setting up the queues between local exchange producers and consumers.

While remote shuffles work with serialized data, local exchanges pass in-memory vectors between threads. This is also the first time we encounter the notion of using columnar encodings to accelerate vectorized execution. Velox became known by its extensive use of such techniques, which we call compressed execution. In this instance, we use dictionaries to slice vectors across multiple destinations – we will discuss it next.

Dictionary Magic​

Query execution often requires changes to the cardinality (number of rows in a result) of the data. This is essentially what filters and joins do – they either reduce the cardinality of the data, e.g., filters or selective joins, or increase the cardinality of the data, e.g. joins with multiple key matches.

Repartitioning in LocalPartition assigns a destination to each row of the input based on the partitioning key. It then makes a vector for each destination with just the rows for that destination. Suppose rows 2, 8 and 100 of the current input hash to 1. Then the vector for destination 1 would only have rows 2, 8 and 100 from the original input. We could make a vector of three rows by copying the data. Instead, we save the copy by wrapping each column of the original input in a DictionaryVector with length 3 and indices 2, 8 and 100. This is much more efficient than copying, especially for wide and nested data.

Later on, the LocalExchange consumer thread running destination 1 will see a DictionaryVector of 3 rows. When this is accessed by the HashAggregation Operator, the aggregation sees a dictionary and will then take the indirection and will access for row 0 the value at 2 in the base (inner) vector, for row 1 the value at 8, and so forth. The consumer for destination 0 does the exact same thing but will access, for example, rows 4, 9 and 50.

The base of the dictionaries is the same on all the consumer threads. Each consumer thread just looks at a different subset. The cores read the same cache lines, but because the base is not written to (read-only), there is no cache-coherence overhead. 

To summarize, a DictionaryVector<T> is a wrapper around any vector of T. The DictionaryVector specifies indices, which give indices into the base vector. Dictionary encoding is typically used when there are few distinct values in a column. Take the strings “experiment” and “baseline”. If a column only has these values, it is far more efficient to represent it as a vector with “experiment” at 0 and “baseline” at 1, and a DictionaryVector that has, say, 1000 elements, where these are either the index 0 or 1.  

Besides this, DictionaryVectors can also be used to denote a subset or a reordering of elements of the base. Because all places that accept vectors also accept DictionaryVectors, the DictionaryVector becomes the universal way of representing selection. This is a central precept of Velox and other modern vectorized engines. We will often come across this concept.

Aggregation and Pipeline Barrier​

We have now arrived at the second pipeline of stage 2. It begins with LocalExchange, which then feeds into HashAggregation. The LocalExchange picks a fraction of the Task's input specific to its local destination, about 1/number-of-drivers of the task's input.

We will talk about hash tables, their specific layout and other tricks in a later post. For now, we look at HashAggregation as a black box. This specific aggregation is a final aggregation, which is a full pipeline barrier that only produces output after all input is received.

How does the aggregation know it has received all its input? Let's trace the progress of the completion signal through the shuffles. A leaf worker knows that it is at end if the Task has received the “no more splits” message in the last task update from the coordinator.

So, if one DataSource inside a tableScan is at end and there are no more splits, this particular TableScan is not blocked and it is at end. This will have the Driver call PartitionedOutput::noMoreInput() on the tableScan's PartitionedOutput. This will cause any buffered data for any destination to get flushed and moved over to OutputBufferManager with a note that no more is coming. OutputBufferManager knows how many Drivers there are in the TableScan pipeline. After it has received this many “no more data coming” messages, it can tell all destinations that this Task will generate no more data for them. 

Now, when stage 2 tasks query stage 1 producers, they will know that they are at end when all the producers have signalled that there is no more data coming. The response to the get data for destination request has a flag identifying the last batch. The ExchangeSource on the stage 2 Task set the no-more-data flag. This is then queried by all the Drivers and each of the Exchange Operators sees this. This then calls noMoreInput in the LocalPartition. This queues up a “no more data” signal in the local exchange queues. If the LocalExchange at the start of the second pipeline of stage 2 sees a “no more data” from each of its sources, then it is at end and noMoreInput is called on the HashAggregation. 

This is how the end of data propagates. Up until now, HashAggregation has produced no output, since the counts are not known until all input is received. Now, HashAggregation starts producing batches of output, which contain the l_partkey value and the number of times this has been seen. This reaches the last PartitionedOutput, which in this case has only one destination, the final worker that produces the result set. This will be at end when all the 100 sources have reported their own end of data.

Recap​

We have finally walked through the distributed execution of a simple query. We presented how data is partitioned between workers in the cluster, and then a second time over inside each worker.

Velox and Presto are designed to aggressively parallelize execution, which means creating distinct, non-overlapping sets of data to process on each thread. The more threads, the more throughput. Also remember that for CPU threads to be effective, they must process tasks that are large enough (often over 100 microseconds worth of cpu time), and not communicate too much with other threads or write to memory other threads are writing to. This is accomplished with local exchanges.

Another important thing to remember from this article is how columnar encodings (DictionaryVectors, in particular) can be used as a zero-copy way of representing selection/reorder/duplication. We will see this pattern again with filters, joins, and other relational operations.

Next we will be looking at joins, filters, and hash aggregations. Stay tuned!