Apache Cassandra - Harry

Clone this repo:
  1. 5fa48c2 Improvements: by Alex Petrov · 3 months ago trunk
  2. 1c7fbc6 Numerious minor improvements while preparing for fuzz-testing 4.0 in earnest: by Alex Petrov · 4 months ago
  3. 42b4fd0 Removed dangling reference to the harry-runner module. by Gianluca Righetto · 7 months ago
  4. ca68b85 Removed dangling reference to the harry-runner module. by Gianluca Righetto · 7 months ago
  5. 41324f5 Removed references to harry-runner from Dockerfile. by Gianluca Righetto · 8 months ago

Harry, a fuzz testing tool for Apache Cassandra

Project aims to generate reproducible workloads that are as close to real-life as possible, while being able to efficiently verify the cluster state against the model without pausing the workload itself.

Reproducibility is achieved by using the PCG family of random number generators and generating schema, configuration, and every step of the workload from the repeatable sequence of random numbers. Schema and configuration are generated from the seed. Each operation is assigned its own monotonically increasing logical timestamp, which preserves logical operation order between different runs.

Efficiency is achieved by employing the features of the PCG random number generator (walking the sequence of random numbers back and forth), and writing value generators in a way that preserves properties of the descriptor it was generated from.

Given a long descriptor can be inflated into some value:

  • value can be deflated back to the descriptor it was generated from (in other words, generation is invertible)
  • two inflated values will sort the same way as two descriptors they were generated from (in other words, generation is order-preserving)

These properties are also preserved for the composite values, such as clustering and partition keys.


Every Harry run starts from Configuration. You can find an example configuration under conf/example.yml.

Clock is a component responsible for mapping logical timestamps to real-time ones. When reproducing test failures, and for validation purposes, a snapshot of such be taken to map a real-time timestamp from the value retrieved from the database to map it back to the logical timestamp of the operation that wrote this value. Given a real-time timestamp, the clock can return a logical timestamp, and vice versa.

Runner is a component that schedules operations that change the cluster (system under test) and model state.

System under test: a Cassandra node or cluster. Default implementation is in_jvm (in-JVM DTest cluster). Harry also supports external clusters.

Model is responsible for tracking logical timestamps that system under test was notified about.

Partition descriptor selector controls how partitions are selected based on the current logical timestamp. The default implementation is a sliding window of partition descriptors that will visit one partition after the other in the window slide_after_repeats times. After that, it will retire one partition descriptor, and pick a new one.

Clustering descriptor selector controls how clustering keys are picked within the partition: how many rows there can be in a partition, how many rows are visited for a logical timestamp, how many operations there will be in batch, what kind of operations there will and how often each kind of operation is going to occur.

Formal Relations Between Entities

To be able to implement efficient models, we had to reduce the amount of state required for validation to a minimum and try to operate on primitive data values most of the time. Any Harry run starts with a seed. Given the same configuration, and the same seed, we're able to make runs deterministic (in other words, records in two clusters created from the same seed are going to have different real-time timestamps, but will be otherwise identical; logical time stamps will also be identical).

Since it‘s clear how to generate things like random schemas, cluster configurations, etc., let’s discuss how we're generating data, and why this type of generation makes validation efficient.

First, we're using PCG family of random number generators, which, besides having nice characteristics that any RNG should have, have two important features:

  • Streams: for single seed, we can have several independent different streams of random numbers.
  • Walkability: PCG generators generate a stream of numbers you can walk back and forth. That is, for any number n that represents a position of the random number in the stream of random numbers, we can get the random number at this position. Conversely, given a random number, we can determine what is its position in the stream. Moreover, knowing a random number, we can determine which number precedes it in the stream of random numbers, and, finally, we can determine how many numbers there are in a stream between the two random numbers.

Out of these operations, determining the next random number in the sequence can be done in constant time, O(1). Advancing generator by n steps can be done in O(log(n)) steps. Since generation is cyclical, advancing the iterator backward is equivalent to advancing it by cardinality - 1 steps. If we're generating 64 bits of entropy, advancing by -1 can be done in 64 steps.

Let's introduce some definitions:

  • lts is a logical timestamp, an entity (number in our case), given by the clock, on which some action occurs
  • m is a modification id, a sequential number of the modification that occurs on lts
  • rts is an approximate real-time as of clock for this run
  • pid is a partition position, a number between 0 and N, for N unique generated partitions
  • pd is a partition descriptor, a unique descriptor identifying the partition
  • cd is a clustering descriptor, a unique descriptor identifying row within some partition

A small introduction that can help to understand the relation between these entities. Hierarchically, the generation process looks as follows:

  • lts is an entry point, from which the decision process starts
  • pd is picked from lts, and determines which partition is going to be visited
  • for (pd, lts) combination, #mods (the number of modification batches) and #rows (the number of rows per modification batch) is determined. m is an index of the modification batch, and i is an index of the operation in the modification batch.
  • cd is picked based on (pd, lts), and n, a sequential number of the operation among all modification batches
  • operation type (whether we're going to perform a write, delete, range delete, etc), columns involved in this operation, and values for the modification are picked depending on the pd, lts, m, and i

Most of this formalization is implemented in OpSelectors, and is relied upon inPartitionVisitor and any implementation of a Model.

Random number generation (see OpSelectors#Rng):

  • rng(i, stream[, e]): returns i'th number drawn from random sequence stream producing values with e bits of entropy (64 bits for now).
  • rng'(s, stream[, e]): returns i of the random number s drawn from the random sequence stream. This function is an inverse of rng.
  • next(rnd, stream[, e]) and prev(rnd, stream[, e]): the next/previous number relative to rnd drawn from random sequence stream.

A simple example of a partition descriptor selector is one that is based on a sliding window of a size s, that slides every n iterations. First, we determine where the window should start for a given lts (in other words, how many times it has already slid). After that, we determine which pd we pick out of s available ones. After picking each one of the s descriptors n times, we retire the oldest descriptor and pick a new one to the window. Window start and offset are then used as input for the rng(start + offset, stream) to make sure descriptors are uniformly distributed.

We can build a clustering descriptor selector in a similar manner. Each partition will use its pd as a stream id, and pick cd from a universe of possible cds of size #cds. On each lts, we pick a random offset, and start picking #ops clusterings from this offset < #cds, and wrap around to index 0 after that. This way, each operation maps to a unique cd, and #op can be determined from cd deterministically.

Data Generation

So far, we have established how to generate partition, clustering, and value descriptors. Now, we need to understand how we can generate data modification statements out of these descriptors in a way that helps us to validate data later.

Since every run has a predefined schema, and by the time we visit a partition we have a logical timestamp, we can make the rest of the decisions: pick a number of batches we‘re about to perform, determine what kind of operations each one of the batches is going to contain, which rows we’re going to visit (clustering for each modification operation).

To generate a write, we need to know which partition we‘re going to visit (in other words, partition descriptor), which row we’d like to modify (in other words, clustering descriptor), which columns we‘re modifying (in other words, a column mask), and, for each modified column - its value. By the time we’re ready to make an actual query to the database, we already know pd, cd, rts, and vds[], which is all we need to “inflate” a write.

To inflate each value descriptor, we take a generator for its datatype, and turn its descriptor into the object. This generation process has the following important properties:

  • it is invertible: for every inflate(vd) -> value, there's deflate(value) -> vd
  • it is order-preserving: compare(vd1, vd2) == compare(inflate(vd1), inflate(vd2))

Inflating pd and cd is slightly more involved than inflating vds, since partition and clustering keys are often composite. This means that inflate(pd) returns an array of objects, rather just a single object: inflate(pd) -> value[], and deflate(value[]) -> pd. Just like inflating value descriptors, inflating keys preserves order.

It is easy to see that, given two modifications: Update(pd1, cd1, [vd1_1, vd2_1, vd3_1], lts1) and Update(pd1, cd1, [vd1_2, vd3_2], lts2), we will end up with a resultset that contains effects of both operations: ResultSetRow(pd1, cd1, [vd1_2@rts2, vd2_1@rts1, vd3_2@rts2]).


Model in Harry ties the rest of the components together and allows us to check whether or not data returned by the cluster actually makes sense. The model relies on the clock, since we have to convert real-time timestamps of the returned values back to logical timestamps, and on descriptor selectors to pick the right partition and rows.

Visible Rows Checker

Let‘s try to put it all together and build a simple model. The simplest one is a visible row checker. It can check if any row in the response returned from the database could have been produced by one of the operations. However, it won’t be able to find errors related to missing rows, and will only notice some cases of erroneously overwritten rows.

In the model, we can see a response from the database in its deflated state. In other words, instead of the actual values returned, we see their descriptors. Every resultset row consists of pd, cd, vds[] (value descriptors), and lts[] (logical timestamps at which these values were written).

To validate, we need to iterate through all operations for this partition, starting with the latest one the model is aware of. This model has no internal state, and validates entire partitions:

void validatePartitionState(long validationLts, List<ResultSetRow> rows) {
  long pd = pdSelector.pd(validationLts, schema);

  for (ResultSetRow row : rows) {
      // iterator that gives us unique lts from the row in descending order
      LongIterator rowLtsIter = descendingIterator(row.lts);
      // iterator that gives us unique lts from the model in descending order
      LongIterator modelLtsIter = descendingIterator(pdSelector, validationLts);

      while (rowLtsIter.hasNext()) {
        long rowLts = rowLtsIter.nextLong();

 	// this model can not check columns whose values were never written or were deleted
        if (rowLts == NO_TIMESTAMP)
          continue outer;

        if (!modelLtsIter.hasNext())
          throw new ValidationException(String.format("Model iterator is exhausted, could not verify %d lts for the row: \n%s %s",
                                                       rowLts, row));

        while (modelLtsIter.hasNext()) {
          long modelLts = modelLtsIter.nextLong();
	  // column was written by the operation that has a lower lts than the current one from the model
          if (modelLts > rowLts)
	  // column was written by the operation that has a higher lts, which contradicts to the model, since otherwise we'd validate it by now
          if (modelLts < rowLts)
              throw new RuntimeException("Can't find a corresponding event id in the model for: " + rowLts + " " + modelLts);

          // Compare values for columns that were supposed to be written with this lts
	  for (int col = 0; col < row.lts.length; col++) {
             if (row.lts[col] != rowLts)

             long m = descriptorSelector.modificationId(pd, row.cd, rowLts, row.vds[col], col);
             long vd = descriptorSelector.vd(pd, row.cd, rowLts, m, col);

	     // If the value model predicts doesn't match the one received from the database, throw an exception
             if (vd != row.vds[col])
                throw new RuntimeException("Returned value doesn't match the model");
	  continue outer;

As you can see, all validation is done using deflated ResultSetRows, which contain enough data to say which logical timestamp each value was written with, and which value descriptor each value has. This model can also validate data concurrently to the ongoing data modification operations.

Quiescent Checker

Let‘s consider one more checker. It’ll be more powerful than the visible rows checker in one way since it can find any inconsistency in data (incorrect timestamp, missing or additional row, rows coming in the wrong order, etc), but it‘ll also have one limitation: it won’t be able to run concurrently with data modification statements. This means that for this model to be used, we should have no in-flight queries, and all queries have to be in a deterministic state by the time we're validating their results.

For this checker, we assume that we have a component that is called Reconciler, which can inflate partition state up to some lts. Reconciler works by simply applying each modification in the same order they were applied to the cluster, and using standard Cassandra data reconciliation rules (last write wins / DELETE wins over INSERT in case of a timestamp collision).

With this component, and knowing that there can be no in-fight queries, we can validate data in the following way:

public void validatePartitionState(Iterator<ResultSetRow> actual, Query query) {
  // find out up the highest completed logical timestamp
  long maxCompleteLts = tracker.maxComplete();

  // get the expected state from reconciler
  Iterator<Reconciler.RowState> expected = reconciler.inflatePartitionState(query.pd, maxCompleteLts, query).iterator(query.reverse);

  // compare actual and expected rows one-by-one in-order
  while (actual.hasNext() && expected.hasNext()) {
    ResultSetRow actualRowState = actual.next();
    Reconciler.RowState expectedRowState = expected.next();

    if (actualRowState.cd != expectedRowState.cd)
      throw new ValidationException("Found a row in the model that is not present in the resultset:\nExpected: %s\nActual: %s",
                                    expectedRowState, actualRowState);

    if (!Arrays.equals(actualRowState.vds, expectedRowState.vds))
      throw new ValidationException("Returned row state doesn't match the one predicted by the model:\nExpected: %s (%s)\nActual:   %s (%s).",
                                    Arrays.toString(expectedRowState.vds), expectedRowState,
                                    Arrays.toString(actualRowState.vds), actualRowState);

    if (!Arrays.equals(actualRowState.lts, expectedRowState.lts))
      throw new ValidationException("Timestamps in the row state don't match ones predicted by the model:\nExpected: %s (%s)\nActual:   %s (%s).",
                                    Arrays.toString(expectedRowState.lts), expectedRowState,
                                    Arrays.toString(actualRowState.lts), actualRowState);

  if (actual.hasNext() || expected.hasNext()) {
    throw new ValidationException("Expected results to have the same number of results, but %s result iterator has more results",
                                   actual.hasNext() ? "actual" : "expected");

If there‘s any mismatch, it’ll be caught right away: if there's an extra row (for example, there were issues in Cassandra that caused it to have duplicate rows), or if some row or even value in the row is missing.

Exhaustive Checker

To be able to both run validation concurrently to modifications and be able to catch all kinds of inconsistencies, we need a more involved checker.

In this checker, we rely on inflating partition state. However, we're most interested in lts, opId, and visibility (whether or not it is still in-flight) of each modification operation. To be able to give a reliable result, we need to make sure we follow these rules:

  • every operation model thinks should be visible, has to be visible
  • every operation model thinks should be invisible, has to be invisible
  • every operation model doesn't know the state of (i.e., it is still in-flight) can be either visible invisible
  • there can be no state in the database that model is not aware of (in other words, we either can explain how a row came to be, or we conclude that the row is erroneous)

A naive way to do this would be to inflate every possible partition state, where every in-flight operation would be either visible or invisible, but this gets costly very quickly since the number of possible combinations grows exponentially. A better (and simpler) way to do this is to iterate all operations and keep the state of “explained” operations:

public class RowValidationState {
  // every column starts in UNOBSERVED, and has to move to either REMOVED, or OBSERVED state
  private final ColumnState[] columnStates;
  // keep track of operations related to each column state
  private final Operation[] causingOperations;

Now, we move through all operations for a given row, starting from the newest ones, towards the oldest ones:

public void validatePartitionState(long verificationLts, PeekingIterator<ResultSetRow> actual_, Query query) {
  // get a list of operations for each cd
  NavigableMap<Long, List<Operation>> operations = inflatePartitionState(query);

  for (Map.Entry<Long, List<Operation>> entry : operations.entrySet()) {
    long cd = entry.getKey();
    List<Operation> ops = entry.getValue();

    // Found a row that is present both in the model and in the resultset
    if (actual.hasNext() && actual.peek().cd == cd) {
      validateRow(new RowValidationState(actual.next), operations);
    } else {
      validateNoRow(cd, operations);

      // Row is not present in the resultset, and we currently look at modifications with a clustering past it
      if (actual.hasNext() && cmp.compare(actual.peek().cd, cd) < 0)
         throw new ValidationException("Couldn't find a corresponding explanation for the row in the model");

  // if there are more rows in the resultset, and we don't have model explanation for them, we've found an issue
  if (actual.hasNext())
    throw new ValidationException("Observed unvalidated rows");

Now, we have to implement validateRow and validateNoRow. validateNoRow is easy: we only need to make sure that a set of operations results in an invisible row. Since we‘re iterating operations in reverse order, if we encounter a delete not followed by any writes, we can conclude that the row is invisible and exit early. If there’s a write that is not followed by a delete, and the row isn‘t covered by a range tombstone, we know it’s an error.

validateRow only has to iterate operations in reverse order until it can explain the value in every column. For example, if a value is UNOBSERVED, and the first thing we encounter is a DELETE that removes this column, we only need to make sure that the value is actually null, in which case we can conclude that the value can be explained as REMOVED.

Similarly, if we encounter an operation that has written the expected value, we conclude that the value is OBSERVED. If there are any seeming inconsistencies between the model and resultset, we have to check whether or not the operation in question is still in flight. If it is, its results may still not be visible, so we can‘t reliably say it’s an error.

To summarize, in order for us to implement an exhaustive checker, we have to iterate operations for each of the rows present in the model in reverse order until we either detect inconsistency that can't be explained by an in-flight operation or until we explain every value in the row.


As you can see, all checkers up till now are almost entirely stateless. Exhaustive and quiescent models rely on DataTracker component that is aware of the in-flight and completed lts, but don't need any other state apart from that, since we can always inflate a complete partition from scratch every time we validate.

While not relying on the state is a useful feature, at least some state is useful to have. For example, if we're validating just a few rows in the partition, right now we have to iterate through each and every lts that has visited this partition and filter out only modifications that have visited it. However, since the model is notified of each started, and, later, finished modification via recordEvent, we can keep track of pd -> (cd -> lts) map. You can check out VisibleRowsChecker as an example of that.


To use Harry, you first need to build a Cassandra in-JVM dtest jar. At the moment of writing, there‘s no official repository where these jars are released, so you’ll have to build it manually:

git clone git@github.com:apache/cassandra.git
cd cassandra
./build-shaded-dtest-jar.sh 4.0-beta4 4.0.0-SNAPSHOT
cd ~/../harry/
mvn package -DskipTests -nsu
mvn dependency:copy-dependencies
java -cp harry-runner/target/harry-runner-0.0.1-SNAPSHOT.jar:$(find harrry-runner/target/dependency/*.jar | tr -s '\n' ':'). harry.runner.HarryRunner

4.0-beta3 is a version of Cassandra which you can find in build.xml, and 4.0.0-SNAPSHOT is a version of dtest jar that'll be installed under org.apache.cassandra:cassandra-dtest-local in your ~/.m2/repository.

Alternatively, you can use a Docker container. For that just run:

git clone git@github.com:apache/cassandra.git
cd cassandra
./build-shaded-dtest-jar.sh 4.0-beta4 4.0.0-SNAPSHOT
cd ~/../harry/
make run

For best effect, uncomment scheduleCorruption(run, executor); in HarryRunner, which will corrupt data in some way so that you could see how Harry detects this corruption.

Each Harry failure contains a complete cluster state, operation log, failure description, and a run configuration. Most of the time, you‘ll be able to just load up the existing cluster state with harry.runner.Reproduce class, which pick up a run configuration from shared/run.yml, and see the same error you’ve seen in your failure log.

When reproducing, make sure to point system_under_test/root in the yaml file to the dump, which is something like ~/harry/shared/cluster-state/1599155256261, and make sure to point validation to the same LTS as the failed one with run.validator.validatePartition(...L).

Because of how our corruptor works, some errors are only reproducible on a specific lts, since they're kind of writing the data “from the future”, so you should also make sure you set the following values from the corresponding values in failure.dump.

    max_seen_lts: your_value
    max_complete_lts: your_value

What's missing

Harry is by no means feature-complete. Main things that are missing are:

  • Some types (such as collections) are not deflatable
  • Some types are implemented, but are not hooked up (blob and text) to DSL/generator
  • Partition deletions are not implemented
  • 2i queries are not implemented
  • Compact storage is not implemented
  • Static columns are not implemented
  • Fault injection is not implemented
  • Runner and scheduler are rather rudimentary and require significant rework and proper scheduling
  • TTL is not supported
  • Some SELECT queries are not supported: LIMIT, IN, GROUP BY, token range queries
  • Pagination is not implemented

Some things, even though are implemented, can be improved or optimized:

  • RNG should be able to yield less than 64 bits of entropy per step
  • State tracking should be done in a compact off-heap data stucture
  • Inflated partition state and per-row operation log should be done in a compact off-heap data structure
  • Exhaustive checker can be significantly optimized
  • Harry shouldn't rely on java-driver for query generation
  • Exhaustive checker should use more precise information from data tracker, not just watermarks
  • Decision-making about when we visit partitions and/or rows should be improved

This list of improvements is incomplete, and should only give the reader a rough idea about the state of the project. Main goal for the initial release was to make it useful, now we can make it fast and feature-complete!


Special thanks to Aleksey Yeschenko, Sam Tunnicliffe, Marcus Eriksson, and Scott Andreas.


Licensed under the Apache License, Version 2.0 (the “License”); you may not use this file except in compliance with the License. You may obtain a copy of the License at


Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an “AS IS” BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.