read_repair_test.py

from contextlib import contextmanager
import glob
import os
import time
from distutils.version import LooseVersion

import pytest
import logging
import subprocess
from uuid import uuid4

from cassandra import ConsistencyLevel, WriteTimeout, ReadTimeout
from cassandra.query import SimpleStatement
from ccmlib.node import Node
from pytest import raises

from dtest import Tester, create_ks, mk_bman_path
from tools.assertions import assert_one
from tools.data import rows_to_list
from tools.jmxutils import JolokiaAgent, make_mbean
from tools.misc import retry_till_success

since = pytest.mark.since
ported_to_in_jvm = pytest.mark.ported_to_in_jvm
logger = logging.getLogger(__name__)

def byteman_validate(node, script, verbose=False, opts=None):
    opts = opts or []
    cdir = node.get_install_dir()
    byteman_cmd = []
    byteman_cmd.append(os.path.join(os.environ['JAVA_HOME'],
                                    'bin',
                                    'java'))
    byteman_cmd.append('-cp')
    jars = [
        glob.glob(os.path.join(cdir, 'build', 'lib', 'jars', 'byteman-[0-9]*.jar'))[0],
        os.path.join(cdir, 'build', '*'),
    ]

    if os.path.exists(os.path.join(cdir, 'modules', 'accord')):
        jars.append(glob.glob(os.path.join(cdir, 'modules', 'accord', 'accord-core', 'build', 'libs', 'accord-core-[0-9].[0-9]-SNAPSHOT.jar'))[0])

    byteman_cmd.append(':'.join(jars))
    byteman_cmd.append('org.jboss.byteman.check.TestScript')
    byteman_cmd.append('-p')
    byteman_cmd.append(node.byteman_port)
    if verbose and '-v' not in opts:
        byteman_cmd.append('-v')
    byteman_cmd.append(script)
    # process = subprocess.Popen(byteman_cmd)
    # out, err = process.communicate()
    out = subprocess.check_output(byteman_cmd)
    if (out is not None) and isinstance(out, bytes):
        out = out.decode()

    has_errors = 'ERROR' in out
    if verbose and not has_errors:
        print (out)

    assert not has_errors, "byteman script didn't compile\n" + out

def build_byteman_submit_command(node, opts):
    cdir = node.get_install_dir()
    byteman_cmd = [os.path.join(os.environ['JAVA_HOME'], 'bin', 'java'),
                   '-cp',
                   glob.glob(os.path.join(cdir, 'build', 'lib', 'jars', 'byteman-submit-[0-9]*.jar'))[0],
                   'org.jboss.byteman.agent.submit.Submit',
                   '-p', node.byteman_port,
                   '-h', node.network_interfaces['binary'][0]] + opts
    return byteman_cmd

def request_verb_timing(node):
    # -y is byteman's built-in flag for reading system props
    byteman_cmd = build_byteman_submit_command(node, ['-y'])
    out = subprocess.check_output(byteman_cmd)
    if (out is not None) and isinstance(out, bytes):
        out = out.decode()
    lines = out.splitlines()
    props = {}
    for line in lines:
        # look for the custom separators, otherwise skip
        if "=" in line and "|" in line:
            key, value = line.split("=")
            split_key = key.split("|")
            ip = split_key[-2].replace("/", "")
            verb = split_key[-1]
            props.setdefault(ip, {}).update({verb: int(value)})
    return props

class TestReadRepair(Tester):

    @pytest.fixture(scope='function')
    def fixture_set_cluster_settings(self, fixture_dtest_setup):
        cluster = fixture_dtest_setup.cluster
        cluster.populate(3)
        # disable dynamic snitch to make replica selection deterministic
        # when we use patient_exclusive_cql_connection, CL=1 and RF=n
        cluster.set_configuration_options(values={'hinted_handoff_enabled': False,
                                                  'endpoint_snitch': 'GossipingPropertyFileSnitch',
                                                  'dynamic_snitch': False})
        for node in cluster.nodelist():
            with open(os.path.join(node.get_conf_dir(), 'cassandra-rackdc.properties'), 'w') as snitch_file:
                snitch_file.write("dc=datacenter1" + os.linesep)
                snitch_file.write("rack=rack1" + os.linesep)
                snitch_file.write("prefer_local=true" + os.linesep)

        cluster.start()

    @since('3.0')
    @ported_to_in_jvm('4.0')
    def test_alter_rf_and_run_read_repair(self, fixture_set_cluster_settings):
        """
        @jira_ticket CASSANDRA-10655
        @jira_ticket CASSANDRA-10657

        Test that querying only a subset of all the columns in a row doesn't confuse read-repair to avoid
        the problem described in CASSANDRA-10655.
        """

        # session is only used to setup & do schema modification. Actual data queries are done directly on
        # each node, using an exclusive connection and CL.ONE
        session = self.patient_cql_connection(self.cluster.nodelist()[0])
        initial_replica, non_replicas = self.do_initial_setup(session)

        # Execute a query at CL.ALL on one of the nodes which was *not* the initial replica. It should trigger a
        # read repair and propagate the data to all 3 nodes.
        # Note: result of the read repair contains only the selected column (a), not all columns
        logger.debug("Executing 'SELECT a...' on non-initial replica to trigger read repair " + non_replicas[0].name)
        read_repair_session = self.patient_exclusive_cql_connection(non_replicas[0])
        assert_one(read_repair_session, "SELECT a FROM alter_rf_test.t1 WHERE k=1", [1], cl=ConsistencyLevel.ALL)

        # The read repair should have repaired the replicas, at least partially (see CASSANDRA-10655)
        # verify by querying each replica in turn.
        value_skipping_disabled = True if self.cluster.version() < '3.4' else False
        self.check_data_on_each_replica(expect_fully_repaired=value_skipping_disabled, initial_replica=initial_replica)

        # Now query again at CL.ALL but this time selecting all columns, which should ensure that 'b' also gets repaired
        query = "SELECT * FROM alter_rf_test.t1 WHERE k=1"
        logger.debug("Executing 'SELECT *...' on non-initial replica to trigger read repair " + non_replicas[0].name)
        assert_one(read_repair_session, query, [1, 1, 1], cl=ConsistencyLevel.ALL)

        # Check each replica individually again now that we expect the data to be fully repaired
        self.check_data_on_each_replica(expect_fully_repaired=True, initial_replica=initial_replica)

    @since('2.1', max_version='3.11.x')
    def test_read_repair_chance(self, fixture_set_cluster_settings):
        """
        @jira_ticket CASSANDRA-12368
        """
        # session is only used to setup & do schema modification. Actual data queries are done directly on
        # each node, using an exclusive connection and CL.ONE
        session = self.patient_cql_connection(self.cluster.nodelist()[0])
        initial_replica, non_replicas = self.do_initial_setup(session)

        # To ensure read repairs are triggered, set the table property to 100%
        logger.debug("Setting table read repair chance to 1")
        session.execute("""ALTER TABLE alter_rf_test.t1 WITH read_repair_chance = 1;""")

        # Execute a query at CL.ONE on one of the nodes which was *not* the initial replica. It should trigger a
        # read repair because read_repair_chance == 1, and propagate the data to all 3 nodes.
        # Note: result of the read repair contains only the selected column (a), not all columns, so we won't expect
        # 'b' to have been fully repaired afterwards.
        logger.debug("Executing 'SELECT a...' on non-initial replica to trigger read repair " + non_replicas[0].name)
        read_repair_session = self.patient_exclusive_cql_connection(non_replicas[0])
        read_repair_session.execute(SimpleStatement("SELECT a FROM alter_rf_test.t1 WHERE k=1",
                                                    consistency_level=ConsistencyLevel.ONE))

        # Query each replica individually to ensure that read repair was triggered. We should expect that only
        # the initial replica has data for both the 'a' and 'b' columns. If the cluster is on > 3.4, the read repair
        # should only have affected the selected column (CASSANDRA-10655), so the other two replicas should only have
        # that data.
        # Note: we need to temporarily set read_repair_chance to 0 while we perform this check.
        logger.debug("Setting table read repair chance to 0 while we verify each replica's data")
        session.execute("""ALTER TABLE alter_rf_test.t1 WITH read_repair_chance = 0;""")
        # The read repair is run in the background, so we spin while checking that the repair has completed
        value_skipping_disabled = True if self.cluster.version() < '3.4' else False
        retry_till_success(self.check_data_on_each_replica,
                           expect_fully_repaired=value_skipping_disabled,
                           initial_replica=initial_replica,
                           timeout=30,
                           bypassed_exception=NotRepairedException)

        # Re-enable global read repair and perform another query on a non-replica. This time the query selects all
        # columns so we also expect the value for 'b' to be repaired.
        logger.debug("Setting table read repair chance to 1")
        session.execute("""ALTER TABLE alter_rf_test.t1 WITH read_repair_chance = 1;""")
        logger.debug("Executing 'SELECT *...' on non-initial replica to trigger read repair " + non_replicas[0].name)
        read_repair_session = self.patient_exclusive_cql_connection(non_replicas[0])
        read_repair_session.execute(SimpleStatement("SELECT * FROM alter_rf_test.t1 WHERE k=1",
                                                    consistency_level=ConsistencyLevel.ONE))

        # Query each replica again to ensure that second read repair was triggered. This time, we expect the
        # data to be fully repaired (both 'a' and 'b' columns) by virtue of the query being 'SELECT *...'
        # As before, we turn off read repair before doing this check.
        logger.debug("Setting table read repair chance to 0 while we verify each replica's data")
        session.execute("""ALTER TABLE alter_rf_test.t1 WITH read_repair_chance = 0;""")
        retry_till_success(self.check_data_on_each_replica,
                           expect_fully_repaired=True,
                           initial_replica=initial_replica,
                           timeout=30,
                           bypassed_exception=NotRepairedException)

    def do_initial_setup(self, session):
        """
        Create a keyspace with rf=1 and a table containing a single row with 2 non-primary key columns.
        Insert 1 row, placing the data on a single initial replica. Then, alter the keyspace to rf=3, but don't
        repair. Tests will execute various reads on the replicas and assert the effects of read repair.
        :param session: Used to perform the schema setup & insert the data
        :return: a tuple containing the node which initially acts as the replica, and a list of the other two nodes
        """
        # Disable speculative retry and [dclocal]read_repair in initial setup.
        session.execute("""CREATE KEYSPACE alter_rf_test
                           WITH replication = {'class': 'SimpleStrategy', 'replication_factor': 1};""")

        options = "speculative_retry='NONE'";
        if self.cluster.version() < '4.0':
            options = options + " AND read_repair_chance=0 AND dclocal_read_repair_chance=0"
        session.execute("CREATE TABLE alter_rf_test.t1 (k int PRIMARY KEY, a int, b int) WITH " + options)

        session.execute("INSERT INTO alter_rf_test.t1 (k, a, b) VALUES (1, 1, 1);")

        # identify the initial replica and trigger a flush to ensure reads come from sstables
        initial_replica, non_replicas = self.identify_initial_placement()
        logger.debug("At RF=1 replica for data is " + initial_replica.name)
        initial_replica.flush()

        # Just some basic validation.
        # At RF=1, it shouldn't matter which node we query, as the actual data should always come from the
        # initial replica when reading at CL ONE
        for n in self.cluster.nodelist():
            logger.debug("Checking " + n.name)
            session = self.patient_exclusive_cql_connection(n)
            assert_one(session, "SELECT * FROM alter_rf_test.t1 WHERE k=1", [1, 1, 1], cl=ConsistencyLevel.ONE)

        # Alter so RF=n but don't repair, calling tests will execute queries to exercise read repair,
        # either at CL.ALL or after setting read_repair_chance to 100%.
        logger.debug("Changing RF from 1 to 3")
        session.execute("""ALTER KEYSPACE alter_rf_test
                           WITH replication = {'class': 'SimpleStrategy', 'replication_factor': 3};""")

        return initial_replica, non_replicas

    def identify_initial_placement(self):
        """
        Identify which node in the 3 node cluster contains the specific key at the point that the test keyspace has
        rf=1.
        :return: tuple containing the initial replica, plus a list of the other 2 replicas.
        """
        nodes = self.cluster.nodelist()
        out, _, _ = nodes[0].nodetool("getendpoints alter_rf_test t1 1")
        address = out.split('\n')[-2]
        initial_replica = None
        non_replicas = []
        for node in nodes:
            if node.address() == address:
                initial_replica = node
            else:
                non_replicas.append(node)

        assert initial_replica is not None, "Couldn't identify initial replica"

        return initial_replica, non_replicas

    def check_data_on_each_replica(self, expect_fully_repaired, initial_replica):
        """
        Perform a SELECT * query at CL.ONE on each replica in turn. If expect_fully_repaired is True, we verify that
        each replica returns the full row being queried. If not, then we only verify that the 'a' column has been
        repaired.
        """
        stmt = SimpleStatement("SELECT * FROM alter_rf_test.t1 WHERE k=1", consistency_level=ConsistencyLevel.ONE)
        logger.debug("Checking all if read repair has completed on all replicas")
        for n in self.cluster.nodelist():
            logger.debug("Checking {n}, {x}expecting all columns"
                         .format(n=n.name, x="" if expect_fully_repaired or n == initial_replica else "not "))
            session = self.patient_exclusive_cql_connection(n)
            res = rows_to_list(session.execute(stmt))
            logger.debug("Actual result: " + str(res))
            expected = [[1, 1, 1]] if expect_fully_repaired or n == initial_replica else [[1, 1, None]]
            if res != expected:
                raise NotRepairedException()

    @since('2.0')
    @ported_to_in_jvm('4.0')
    def test_range_slice_query_with_tombstones(self, fixture_set_cluster_settings):
        """
        @jira_ticket CASSANDRA-8989
        @jira_ticket CASSANDRA-9502

        Range-slice queries with CL>ONE do unnecessary read-repairs.
        Reading from table which contains collection type using token function and with CL > ONE causes overwhelming writes to replicas.


        It's possible to check the behavior with tracing - pattern matching in system_traces.events.activity
        """
        node1 = self.cluster.nodelist()[0]
        session1 = self.patient_exclusive_cql_connection(node1)

        session1.execute("CREATE KEYSPACE ks WITH replication = {'class': 'NetworkTopologyStrategy', 'datacenter1': 2}")
        session1.execute("""
            CREATE TABLE ks.cf (
                key    int primary key,
                value  double,
                txt    text
            );
        """)

        for n in range(1, 2500):
            str = "foo bar %d iuhiu iuhiu ihi" % n
            session1.execute("INSERT INTO ks.cf (key, value, txt) VALUES (%d, %d, '%s')" % (n, n, str))

        self.cluster.flush()
        self.cluster.stop()
        self.cluster.start()
        session1 = self.patient_exclusive_cql_connection(node1)

        for n in range(1, 1000):
            session1.execute("DELETE FROM ks.cf WHERE key = %d" % (n))

        time.sleep(1)

        node1.flush()

        time.sleep(1)

        query = SimpleStatement("SELECT * FROM ks.cf LIMIT 100", consistency_level=ConsistencyLevel.LOCAL_QUORUM)
        future = session1.execute_async(query, trace=True)
        future.result()
        trace = future.get_query_trace(max_wait=120)
        self.pprint_trace(trace)
        for trace_event in trace.events:
            # Step 1, find coordinator node:
            activity = trace_event.description
            assert "Appending to commitlog" not in activity
            assert "Adding to cf memtable" not in activity
            assert "Acquiring switchLock read lock" not in activity

    @since('3.0')
    @ported_to_in_jvm('4.0')
    def test_gcable_tombstone_resurrection_on_range_slice_query(self, fixture_set_cluster_settings):
        """
        @jira_ticket CASSANDRA-11427

        Range queries before the 11427 will trigger read repairs for puregable tombstones on hosts that already compacted given tombstones.
        This will result in constant transfer and compaction actions sourced by few nodes seeding purgeable tombstones and triggered e.g.
        by periodical jobs scanning data range wise.
        """

        node1, node2, _ = self.cluster.nodelist()

        session1 = self.patient_cql_connection(node1)
        create_ks(session1, 'gcts', 3)
        query = """
            CREATE TABLE gcts.cf1 (
                key text,
                c1 text,
                PRIMARY KEY (key, c1)
            )
            WITH gc_grace_seconds=0
            AND compaction = {'class': 'SizeTieredCompactionStrategy', 'enabled': 'false'};
        """
        session1.execute(query)

        # create row tombstone
        delete_stmt = SimpleStatement("DELETE FROM gcts.cf1 WHERE key = 'a'", consistency_level=ConsistencyLevel.ALL)
        session1.execute(delete_stmt)

        # flush single sstable with tombstone
        node1.flush()
        node2.flush()

        # purge tombstones from node2 (gc grace 0)
        node2.compact()

        # execute range slice query, which should not trigger read-repair for purged TS
        future = session1.execute_async(SimpleStatement("SELECT * FROM gcts.cf1", consistency_level=ConsistencyLevel.ALL), trace=True)
        future.result()
        trace = future.get_query_trace(max_wait=120)
        self.pprint_trace(trace)
        for trace_event in trace.events:
            activity = trace_event.description
            assert "Sending READ_REPAIR message" not in activity


    @since('3.0')
    def test_tracing_does_not_interfere_with_digest_calculation(self):
        """
        Test that enabling tracing doesn't interfere with digest responses when using RandomPartitioner.
        The use of a threadlocal MessageDigest for generating both DigestResponse messages and for
        calculating tokens meant that the DigestResponse was always incorrect when both RP and tracing
        were enabled, leading to unnecessary data reads.

        @jira_ticket CASSANDRA-13964
        """
        cluster = self.cluster
        cluster.populate(3)
        opts = {'write_request_timeout_in_ms': 30000, 'read_request_timeout_in_ms': 30000}
        if cluster.version() >= LooseVersion('4.1'):
            opts['native_transport_timeout'] = '30s'
        cluster.set_configuration_options(values=opts)
        cluster.set_partitioner("org.apache.cassandra.dht.RandomPartitioner")
        cluster.start(jvm_args=['-Dcassandra.wait_for_tracing_events_timeout_secs=15'])

        node1 = cluster.nodelist()[0]
        session = self.patient_cql_connection(node1)
        create_ks(session, 'ks', 3)

        session.execute("""
            CREATE TABLE ks.users (
                userid uuid PRIMARY KEY,
                firstname text,
                lastname text,
                age int
            );
        """)

        insert = session.prepare(
              "INSERT INTO ks.users (userid, firstname, lastname, age) "
              "VALUES (?, 'Frodo', 'Baggins', 32)")
        insert.consistency_level = ConsistencyLevel.ALL

        select = session.prepare(
              "SELECT firstname, lastname "
              "FROM ks.users WHERE userid = ?")
        select.consistency_level = ConsistencyLevel.ALL

        for _ in range(10):
            id = uuid4()
            session.execute(insert.bind((id,)), timeout=30)
            res = session.execute(select.bind((id,)), timeout=30, trace=True)
            assert 1 == len(res.response_future.get_query_trace_ids())

        rr_count = make_mbean('metrics', type='ReadRepair', name='RepairedBlocking')
        with JolokiaAgent(node1) as jmx:
            # the MBean may not have been initialized, in which case Jolokia agent will return
            # a HTTP 404 response. If we receive such, we know that no digest mismatch was reported
            # If we are able to read the MBean attribute, assert that the count is 0
            if jmx.has_mbean(rr_count):
                # expect 0 digest mismatches
                assert 0 == jmx.read_attribute(rr_count, 'Count')
            else:
                pass

    def pprint_trace(self, trace):
        """Pretty print a trace"""
        if logging.root.level == logging.DEBUG:
            print(("-" * 40))
            for t in trace.events:
                print(("%s\t%s\t%s\t%s" % (t.source, t.source_elapsed, t.description, t.thread_name)))
            print(("-" * 40))


def quorum(query_string):
    return SimpleStatement(query_string=query_string, consistency_level=ConsistencyLevel.QUORUM)


kcv = lambda k, c, v: [k, c, v]


listify = lambda results: [list(r) for r in results]


class StorageProxy(object):

    def __init__(self, node):
        assert isinstance(node, Node)
        self.node = node
        self.jmx = JolokiaAgent(node)

    def start(self):
        self.jmx.start()

    def stop(self):
        self.jmx.stop()

    def _get_metric(self, metric):
        mbean = make_mbean("metrics", type="ReadRepair", name=metric)
        return self.jmx.read_attribute(mbean, "Count")

    @property
    def blocking_read_repair(self):
        return self._get_metric("RepairedBlocking")

    @property
    def speculated_rr_read(self):
        return self._get_metric("SpeculatedRead")

    @property
    def speculated_rr_write(self):
        return self._get_metric("SpeculatedWrite")

    def get_table_metric(self, keyspace, table, metric, attr="Count"):
        mbean = make_mbean("metrics", keyspace=keyspace, scope=table, type="Table", name=metric)
        return self.jmx.read_attribute(mbean, attr)

    def __enter__(self):
        """ For contextmanager-style usage. """
        self.start()
        return self

    def __exit__(self, exc_type, value, traceback):
        """ For contextmanager-style usage. """
        self.stop()


class TestSpeculativeReadRepair(Tester):

    @pytest.fixture(scope='function', autouse=True)
    def fixture_set_cluster_settings(self, fixture_dtest_setup):
        cluster = fixture_dtest_setup.cluster
        cluster.set_configuration_options(values={'hinted_handoff_enabled': False,
                                                  'dynamic_snitch': False,
                                                  'write_request_timeout_in_ms': 1000,
                                                  'read_request_timeout_in_ms': 1000})
        cluster.populate(3, install_byteman=True, debug=True)
        byteman_validate(cluster.nodelist()[0], mk_bman_path('read_repair/sorted_live_endpoints.btm'), verbose=True)
        cluster.start(jvm_args=['-XX:-PerfDisableSharedMem'])
        session = fixture_dtest_setup.patient_exclusive_cql_connection(cluster.nodelist()[0], timeout=2)

        session.execute("CREATE KEYSPACE ks WITH replication = {'class': 'SimpleStrategy', 'replication_factor': 3}")
        session.execute("CREATE TABLE ks.tbl (k int, c int, v int, primary key (k, c)) WITH speculative_retry = '400ms';")

    def get_cql_connection(self, node, **kwargs):
        return self.patient_exclusive_cql_connection(node, retry_policy=None, **kwargs)

    @since('4.0')
    def test_failed_read_repair(self):
        """
        If none of the disagreeing nodes ack the repair mutation, the read should fail
        """
        node1, node2, node3 = self.cluster.nodelist()
        assert isinstance(node1, Node)
        assert isinstance(node2, Node)
        assert isinstance(node3, Node)

        session = self.get_cql_connection(node1, timeout=2)
        session.execute(quorum("INSERT INTO ks.tbl (k, c, v) VALUES (1, 0, 1)"))

        node2.byteman_submit([mk_bman_path('read_repair/stop_writes.btm')])
        node3.byteman_submit([mk_bman_path('read_repair/stop_writes.btm')])
        script_version = '_5_1' if self.cluster.version() >= LooseVersion('5.1') else ''
        node2.byteman_submit([mk_bman_path('read_repair/stop_rr_writes{}.btm'.format(script_version))])
        node3.byteman_submit([mk_bman_path('read_repair/stop_rr_writes{}.btm'.format(script_version))])

        with raises(WriteTimeout):
            session.execute(quorum("INSERT INTO ks.tbl (k, c, v) VALUES (1, 1, 2)"))

        node2.byteman_submit([mk_bman_path('read_repair/sorted_live_endpoints.btm')])
        session = self.get_cql_connection(node2)
        with StorageProxy(node2) as storage_proxy:
            assert storage_proxy.blocking_read_repair == 0
            assert storage_proxy.speculated_rr_read == 0
            assert storage_proxy.speculated_rr_write == 0

            with raises(ReadTimeout):
                session.execute(quorum("SELECT * FROM ks.tbl WHERE k=1"))

            assert storage_proxy.blocking_read_repair > 0
            assert storage_proxy.speculated_rr_read == 0
            assert storage_proxy.speculated_rr_write > 0

    @since('4.0')
    def test_normal_read_repair(self):
        """ test the normal case """
        node1, node2, node3 = self.cluster.nodelist()
        assert isinstance(node1, Node)
        assert isinstance(node2, Node)
        assert isinstance(node3, Node)
        session = self.get_cql_connection(node1, timeout=2)

        session.execute(quorum("INSERT INTO ks.tbl (k, c, v) VALUES (1, 0, 1)"))

        node2.byteman_submit([mk_bman_path('read_repair/stop_writes.btm')])
        node3.byteman_submit([mk_bman_path('read_repair/stop_writes.btm')])

        session.execute("INSERT INTO ks.tbl (k, c, v) VALUES (1, 1, 2)")

        # re-enable writes
        node2.byteman_submit(['-u', mk_bman_path('read_repair/stop_writes.btm')])

        node2.byteman_submit([mk_bman_path('read_repair/sorted_live_endpoints.btm')])
        coordinator = node2
        # Stop reads on coordinator in order to make sure we do not go through
        # the messaging service for the local reads
        with StorageProxy(node2) as storage_proxy, stop_reads(coordinator):
            assert storage_proxy.blocking_read_repair == 0
            assert storage_proxy.speculated_rr_read == 0
            assert storage_proxy.speculated_rr_write == 0

            session = self.get_cql_connection(coordinator)
            expected = [kcv(1, 0, 1), kcv(1, 1, 2)]
            results = session.execute(quorum("SELECT * FROM ks.tbl WHERE k=1"))
            assert listify(results) == expected

            assert storage_proxy.blocking_read_repair == 1
            assert storage_proxy.speculated_rr_read == 0
            assert storage_proxy.speculated_rr_write == 0

    @since('4.0')
    def test_speculative_data_request(self):
        """ If one node doesn't respond to a full data request, it should query the other """
        node1, node2, node3 = self.cluster.nodelist()
        assert isinstance(node1, Node)
        assert isinstance(node2, Node)
        assert isinstance(node3, Node)
        session = self.get_cql_connection(node1, timeout=2)

        session.execute(quorum("INSERT INTO ks.tbl (k, c, v) VALUES (1, 0, 1)"))

        node2.byteman_submit([mk_bman_path('read_repair/stop_writes.btm')])
        node3.byteman_submit([mk_bman_path('read_repair/stop_writes.btm')])

        session.execute("INSERT INTO ks.tbl (k, c, v) VALUES (1, 1, 2)")

        # re-enable writes
        node2.byteman_submit(['-u', mk_bman_path('read_repair/stop_writes.btm')])

        node1.byteman_submit([mk_bman_path('read_repair/sorted_live_endpoints.btm')])
        version = self.cluster.cassandra_version()
        if version < '4.1':
            node1.byteman_submit([mk_bman_path('request_verb_timing.btm')])
        else:
            node1.byteman_submit([mk_bman_path('post4.0/request_verb_timing.btm')])

        with StorageProxy(node1) as storage_proxy:
            assert storage_proxy.blocking_read_repair == 0
            assert storage_proxy.speculated_rr_read == 0
            assert storage_proxy.speculated_rr_write == 0

            session = self.get_cql_connection(node1)
            node2.byteman_submit([mk_bman_path('read_repair/stop_data_reads.btm')])
            results = session.execute(quorum("SELECT * FROM ks.tbl WHERE k=1"))

            timing = request_verb_timing(node1)
            repair_req_node3 = timing[node3.ip_addr].get('READ_REPAIR_REQ')
            repair_req_node2 = timing[node2.ip_addr].get('READ_REPAIR_REQ')
            assert listify(results) == [kcv(1, 0, 1), kcv(1, 1, 2)]

            assert storage_proxy.blocking_read_repair == 1
            assert storage_proxy.speculated_rr_read == 1
            
            # under normal circumstances we don't expect a speculated write here,
            # but the repair request to node 3 may timeout due to CPU contention and
            # then a speculated write is sent to node 2, so we just make sure that the
            # request to node 2 didn't happen before the request to node 3
            assert storage_proxy.speculated_rr_write == 0 or repair_req_node2 > repair_req_node3

    @since('4.0')
    def test_speculative_write(self):
        """ if one node doesn't respond to a read repair mutation, it should be sent to the remaining node """
        node1, node2, node3 = self.cluster.nodelist()
        assert isinstance(node1, Node)
        assert isinstance(node2, Node)
        assert isinstance(node3, Node)
        session = self.get_cql_connection(node1, timeout=2)

        session.execute(quorum("INSERT INTO ks.tbl (k, c, v) VALUES (1, 0, 1)"))

        node2.byteman_submit([mk_bman_path('read_repair/stop_writes.btm')])
        node3.byteman_submit([mk_bman_path('read_repair/stop_writes.btm')])

        session.execute("INSERT INTO ks.tbl (k, c, v) VALUES (1, 1, 2)")

        # re-enable writes on node 3, leave them off on node2
        script_version = '_5_1' if self.cluster.version() >= LooseVersion('5.1') else ''
        node2.byteman_submit([mk_bman_path('read_repair/stop_rr_writes{}.btm'.format(script_version))])

        node1.byteman_submit([mk_bman_path('read_repair/sorted_live_endpoints.btm')])
        with StorageProxy(node1) as storage_proxy:
            assert storage_proxy.blocking_read_repair == 0
            assert storage_proxy.speculated_rr_read == 0
            assert storage_proxy.speculated_rr_write == 0

            session = self.get_cql_connection(node1)
            expected = [kcv(1, 0, 1), kcv(1, 1, 2)]
            results = session.execute(quorum("SELECT * FROM ks.tbl WHERE k=1"))
            assert listify(results) == expected

            assert storage_proxy.blocking_read_repair == 1
            assert storage_proxy.speculated_rr_read == 0
            assert storage_proxy.speculated_rr_write == 1

    @since('4.0')
    def test_quorum_requirement(self):
        """
        Even if we speculate on every stage, we should still only require a quorum of responses for success
        """
        node1, node2, node3 = self.cluster.nodelist()
        assert isinstance(node1, Node)
        assert isinstance(node2, Node)
        assert isinstance(node3, Node)
        session = self.get_cql_connection(node1, timeout=2)

        session.execute(quorum("INSERT INTO ks.tbl (k, c, v) VALUES (1, 0, 1)"))

        node2.byteman_submit([mk_bman_path('read_repair/stop_writes.btm')])
        node3.byteman_submit([mk_bman_path('read_repair/stop_writes.btm')])

        session.execute("INSERT INTO ks.tbl (k, c, v) VALUES (1, 1, 2)")

        # re-enable writes
        node2.byteman_submit(['-u', mk_bman_path('read_repair/stop_writes.btm')])
        node3.byteman_submit(['-u', mk_bman_path('read_repair/stop_writes.btm')])

        # force endpoint order
        node1.byteman_submit([mk_bman_path('read_repair/sorted_live_endpoints.btm')])

        node2.byteman_submit([mk_bman_path('read_repair/stop_data_reads.btm')])
        script_version = '_5_1' if self.cluster.version() >= LooseVersion('5.1') else ''
        node3.byteman_submit([mk_bman_path('read_repair/stop_rr_writes{}.btm'.format(script_version))])

        with StorageProxy(node1) as storage_proxy:
            assert storage_proxy.get_table_metric("ks", "tbl", "SpeculativeRetries") == 0
            assert storage_proxy.blocking_read_repair == 0
            assert storage_proxy.speculated_rr_read == 0
            assert storage_proxy.speculated_rr_write == 0

            session = self.get_cql_connection(node1)
            expected = [kcv(1, 0, 1), kcv(1, 1, 2)]
            results = session.execute(quorum("SELECT * FROM ks.tbl WHERE k=1"))
            assert listify(results) == expected

            assert storage_proxy.get_table_metric("ks", "tbl", "SpeculativeRetries") == 0
            assert storage_proxy.blocking_read_repair == 1
            assert storage_proxy.speculated_rr_read == 1
            assert storage_proxy.speculated_rr_write == 1

    @since('4.0')
    def test_quorum_requirement_on_speculated_read(self):
        """
        Even if we speculate on every stage, we should still only require a quorum of responses for success
        """
        node1, node2, node3 = self.cluster.nodelist()
        assert isinstance(node1, Node)
        assert isinstance(node2, Node)
        assert isinstance(node3, Node)
        session = self.get_cql_connection(node1, timeout=2)

        session.execute(quorum("INSERT INTO ks.tbl (k, c, v) VALUES (1, 0, 1)"))

        node2.byteman_submit([mk_bman_path('read_repair/stop_writes.btm')])
        node3.byteman_submit([mk_bman_path('read_repair/stop_writes.btm')])

        session.execute("INSERT INTO ks.tbl (k, c, v) VALUES (1, 1, 2)")

        # re-enable writes
        node2.byteman_submit(['-u', mk_bman_path('read_repair/stop_writes.btm')])
        node3.byteman_submit(['-u', mk_bman_path('read_repair/stop_writes.btm')])

        # force endpoint order
        node1.byteman_submit([mk_bman_path('read_repair/sorted_live_endpoints.btm')])

        node2.byteman_submit([mk_bman_path('read_repair/stop_digest_reads.btm')])
        node3.byteman_submit([mk_bman_path('read_repair/stop_data_reads.btm')])
        script_version = '_5_1' if self.cluster.version() >= LooseVersion('5.1') else ''
        node2.byteman_submit([mk_bman_path('read_repair/stop_rr_writes{}.btm'.format(script_version))])

        with StorageProxy(node1) as storage_proxy:
            assert storage_proxy.get_table_metric("ks", "tbl", "SpeculativeRetries") == 0
            assert storage_proxy.blocking_read_repair == 0
            assert storage_proxy.speculated_rr_read == 0
            assert storage_proxy.speculated_rr_write == 0

            session = self.get_cql_connection(node1)
            expected = [kcv(1, 0, 1), kcv(1, 1, 2)]
            results = session.execute(quorum("SELECT * FROM ks.tbl WHERE k=1"))
            assert listify(results) == expected

            assert storage_proxy.get_table_metric("ks", "tbl", "SpeculativeRetries") == 1
            assert storage_proxy.blocking_read_repair == 1
            assert storage_proxy.speculated_rr_read == 0  # there shouldn't be any replicas to speculate on
            assert storage_proxy.speculated_rr_write == 1


@contextmanager
def _byteman_cycle(nodes, scripts):
    script_path = lambda name: mk_bman_path('read_repair/') + name + '.btm'

    for script in scripts:
        byteman_validate(nodes[0], script_path(script))

    for node in nodes:
        assert isinstance(node, Node)
        for name in scripts:

            print(node.name)
            node.byteman_submit([script_path(name)])
    yield

    for node in nodes:
        for name in scripts:
            print(node.name)
            node.byteman_submit(['-u', script_path(name)])


@contextmanager
def stop_writes(*nodes, kind='all'):
    assert kind in ('all', 'normal', 'repair')
    normal = 'stop_writes'
    repair = 'stop_rr_writes'
    with _byteman_cycle(nodes, {'normal': [normal], 'repair': [repair], 'all': [normal, repair]}[kind]):
        yield


@contextmanager
def stop_reads(*nodes, kind='all'):
    data = 'stop_data_reads'
    digest = 'stop_digest_reads'
    with _byteman_cycle(nodes, {'data': [data], 'digest': [digest], 'all': [data, digest]}[kind]):
        yield

kcvv = lambda k, c, v1, v2: [k, c, v1, v2]


@since('4.0')
@ported_to_in_jvm('4.0')
class TestReadRepairGuarantees(Tester):

    @pytest.fixture(scope='function', autouse=True)
    def fixture_set_cluster_settings(self, fixture_dtest_setup):
        cluster = fixture_dtest_setup.cluster
        cluster.set_configuration_options(values={'hinted_handoff_enabled': False,
                                                  'dynamic_snitch': False,
                                                  'write_request_timeout_in_ms': 500,
                                                  'read_request_timeout_in_ms': 500})
        cluster.populate(3, install_byteman=True, debug=True).start(jvm_args=['-XX:-PerfDisableSharedMem'])
        session = fixture_dtest_setup.patient_exclusive_cql_connection(cluster.nodelist()[0], timeout=2)

        session.execute("CREATE KEYSPACE ks WITH replication = {'class': 'SimpleStrategy', 'replication_factor': 3}")

    def get_cql_connection(self, node, **kwargs):
        return self.patient_exclusive_cql_connection(node, retry_policy=None, **kwargs)

    @pytest.mark.parametrize("repair_type,expect_monotonic",
                             (('blocking', True), ('none', False)),
                             ids=('blocking', 'none'))
    def test_monotonic_reads(self, repair_type, expect_monotonic):
        """ 
        tests how read repair provides, or breaks, read monotonicity
        blocking read repair should maintain monotonic quorum reads, async and none should not
         """
        assert repair_type in ('blocking', 'async', 'none')
        node1, node2, node3 = self.cluster.nodelist()

        session = self.get_cql_connection(node1, timeout=2)
        ddl = "CREATE TABLE ks.tbl (k int, c int, v1 int, v2 int, primary key (k, c)) WITH read_repair = '" + repair_type + "';"
        print (ddl)
        session.execute(ddl)

        session.execute(quorum("INSERT INTO ks.tbl (k, c, v1, v2) VALUES (1, 0, 1, 1)"))

        with stop_writes(node2, node3):
            session.execute("INSERT INTO ks.tbl (k, c, v1, v2) VALUES (1, 0, 2, 2)")

        with stop_reads(node3), stop_writes(node3):

            if expect_monotonic:
                results = session.execute(quorum("SELECT * FROM ks.tbl WHERE k=1"))
            else:
                # if we don't expect monotonicity, read repair writes shouldn't block
                with stop_writes(node2, kind='repair'):
                    results = session.execute(quorum("SELECT * FROM ks.tbl WHERE k=1"))
            assert listify(results) == [kcvv(1, 0, 2, 2)]

        session = self.get_cql_connection(node3, timeout=2)
        with stop_reads(node1):
            results = session.execute(quorum("SELECT * FROM ks.tbl WHERE k=1"))
            if expect_monotonic:
                assert listify(results) == [kcvv(1, 0, 2, 2)]
            else:
                assert listify(results) == [kcvv(1, 0, 1, 1)]


    @pytest.mark.parametrize("repair_type,expect_atomic",
                             (('blocking', False), ('none', True)),
                             ids=('blocking', 'none'))
    def test_atomic_writes(self, repair_type, expect_atomic):
        """ 
        tests how read repair provides, or breaks, write atomicity
        'none' read repair should maintain atomic writes, blocking and async should not
         """
        assert repair_type in ('blocking', 'async', 'none')
        node1, node2, node3 = self.cluster.nodelist()

        session = self.get_cql_connection(node1, timeout=2)
        ddl = "CREATE TABLE ks.tbl (k int, c int, v1 int, v2 int, primary key (k, c)) WITH read_repair = '" + repair_type + "';"
        print (ddl)
        session.execute(ddl)

        session.execute(quorum("INSERT INTO ks.tbl (k, c, v1, v2) VALUES (1, 0, 1, 1)"))

        with stop_writes(node2, node3):
            session.execute("INSERT INTO ks.tbl (k, c, v1, v2) VALUES (1, 0, 2, 2)")

        with stop_reads(node3), stop_writes(node3):
            results = session.execute(quorum("SELECT v1 FROM ks.tbl WHERE k=1"))
            assert listify(results) == [[2]]

        # make sure async read repair has a chance to write the repair value
        if repair_type == 'async':
            time.sleep(1)

        session = self.get_cql_connection(node3, timeout=2)
        with stop_reads(node1):
            results = session.execute(quorum("SELECT * FROM ks.tbl WHERE k=1"))
            if expect_atomic:
                assert listify(results) == [kcvv(1, 0, 1, 1)]
            else:
                assert listify(results) == [kcvv(1, 0, 2, 1)]


class NotRepairedException(Exception):
    """
    Thrown to indicate that the data on a replica hasn't been doesn't match what we'd expect if a
    specific read repair has run. See check_data_on_each_replica.
    """
    pass