| Description | +This ```POST``` method probes the interpreter processes of the given interpreter setting + and reports whether they answer. Useful right after a restart, which only closes the + processes and leaves the next use to start them again. The check never starts an + interpreter: a setting with nothing running is reported with the reason + ```NOT_RUNNING```. A setting can own one process per user or note, so the response holds + an entry per interpreter group. The server bounds how long it waits, and a group whose + probe did not answer in time is reported with the reason ```PROBE_TIMEOUT``` rather than + holding the request. | +
| URL | +```http://[zeppelin-server]:[zeppelin-port]/api/interpreter/setting/[interpreter ID]/healthcheck``` | +
| Success code | +200 | +
| Fail code | +403 if the note of the request may not be acted on, 404 if there is no such interpreter setting | +
| Sample JSON input (Optional) | +The note to authorize the caller against, needed only when the caller is not allowed to + reach the interpreter endpoints on their own. Running a paragraph of that note is enough. + +```json +{ + "noteId": "2AVQJVC8N" +} +``` + | +
| Sample JSON response | +```healthy``` follows whether the interpreter answered. ```alive``` and ```running``` + are left out when no probe was made, so that an interpreter found dead stays + distinguishable from one that was never asked. + +```json +{ + "status": "OK", + "message": "", + "body": { + "settingId": "2CH2VMY7B", + "settingName": "spark", + "groups": [ + { + "groupId": "2CH2VMY7B-shared_process", + "healthy": true, + "reason": "OK", + "alive": true, + "running": true, + "probeTookMs": 3 + }, + { + "groupId": "2CH2VMY7B-user2", + "healthy": false, + "reason": "NOT_REACHABLE", + "alive": true, + "running": false, + "probeTookMs": 1004 + }, + { + "groupId": "2CH2VMY7B-user3", + "healthy": false, + "reason": "PROBE_TIMEOUT", + "probeTookMs": 3000 + } + ] + } +} +``` + | +
| Sample JSON response of an interpreter that is not running | ++ +```json +{ + "status": "OK", + "message": "", + "body": { + "settingId": "2CH2VMY7B", + "settingName": "spark", + "reason": "NOT_RUNNING", + "groups": [] + } +} +``` + | +
A setting owns zero or more {@link ManagedInterpreterGroup}s, one per interpreter process, so + * the result reports an entry per group instead of one verdict for the setting. + * + *
{@code healthy} follows {@code InterpreterClient#isRunning()}, documented as "the interpreter + * can communicate with server", while {@code isAlive()} only states that a process exists. Both are + * reported so that a process which is up but no longer answering stays distinguishable from one + * that is gone. + * + *
{@code alive} and {@code running} are {@code null} whenever no probe was performed, so that
+ * "probed and found dead" does not read the same as "never probed".
+ */
+public class InterpreterHealthCheck {
+
+ /**
+ * Why a group is, or is not, healthy. Only {@link #OK} means healthy; the remaining reasons are
+ * not equally bad, and a client is expected to tell them apart rather than to treat every one of
+ * them as a failure.
+ */
+ public enum Reason {
+ /** The interpreter answered the probe. */
+ OK,
+ /** A process handle exists, but the interpreter did not answer. */
+ NOT_REACHABLE,
+ /** The probe did not finish within the budget of this health check. */
+ PROBE_TIMEOUT,
+ /** A process is currently being launched for this group, which is not a failure. */
+ LAUNCHING,
+ /** No process is running, which is the normal state right after a restart. */
+ NOT_RUNNING
+ }
+
+ private final String settingId;
+ private final String settingName;
+ private final Reason reason;
+ private final List Instances are created through the factory methods so that {@code healthy} is always derived
+ * from the same probe result the reason was derived from, and the two can never disagree.
+ */
+ public static class GroupHealth {
+
+ private final String groupId;
+ private final boolean healthy;
+ private final Reason reason;
+ private final Boolean alive;
+ private final Boolean running;
+ private final long probeTookMs;
+
+ private GroupHealth(String groupId, Reason reason, Boolean alive, Boolean running,
+ long probeTookMs) {
+ this.groupId = groupId;
+ this.reason = reason;
+ this.alive = alive;
+ this.running = running;
+ this.probeTookMs = probeTookMs;
+ this.healthy = Boolean.TRUE.equals(running);
+ }
+
+ /** The probe came back, and whether the interpreter answered decides the reason. */
+ public static GroupHealth probed(String groupId, boolean alive, boolean running,
+ long probeTookMs) {
+ return new GroupHealth(groupId, running ? Reason.OK : Reason.NOT_REACHABLE,
+ alive, running, probeTookMs);
+ }
+
+ /**
+ * The probe did not come back within the budget. Nothing is known about the process, hence no
+ * {@code alive} or {@code running}.
+ */
+ public static GroupHealth timedOut(String groupId, long probeTookMs) {
+ return new GroupHealth(groupId, Reason.PROBE_TIMEOUT, null, null, probeTookMs);
+ }
+
+ /**
+ * The probe itself failed, which says the interpreter could not be reached but leaves the state
+ * of the process unknown, hence no {@code alive} or {@code running}.
+ */
+ public static GroupHealth probeFailed(String groupId, long probeTookMs) {
+ return new GroupHealth(groupId, Reason.NOT_REACHABLE, null, null, probeTookMs);
+ }
+
+ /** A process is being launched for this group, so it is deliberately left unprobed. */
+ public static GroupHealth launching(String groupId) {
+ return new GroupHealth(groupId, Reason.LAUNCHING, null, null, 0);
+ }
+
+ /** The group carries no process handle, so there is nothing to probe. */
+ public static GroupHealth notRunning(String groupId) {
+ return new GroupHealth(groupId, Reason.NOT_RUNNING, null, null, 0);
+ }
+
+ public String getGroupId() {
+ return groupId;
+ }
+
+ public boolean isHealthy() {
+ return healthy;
+ }
+
+ public Reason getReason() {
+ return reason;
+ }
+
+ public Boolean getAlive() {
+ return alive;
+ }
+
+ public Boolean getRunning() {
+ return running;
+ }
+
+ public long getProbeTookMs() {
+ return probeTookMs;
+ }
+ }
+}
diff --git a/zeppelin-server/src/main/java/org/apache/zeppelin/interpreter/InterpreterHealthChecker.java b/zeppelin-server/src/main/java/org/apache/zeppelin/interpreter/InterpreterHealthChecker.java
new file mode 100644
index 00000000000..9dce1d50c10
--- /dev/null
+++ b/zeppelin-server/src/main/java/org/apache/zeppelin/interpreter/InterpreterHealthChecker.java
@@ -0,0 +1,201 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You 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
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * 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.
+ */
+
+package org.apache.zeppelin.interpreter;
+
+import com.google.common.annotations.VisibleForTesting;
+
+import org.apache.zeppelin.interpreter.InterpreterHealthCheck.GroupHealth;
+import org.apache.zeppelin.interpreter.remote.RemoteInterpreterProcess;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+
+import java.util.ArrayList;
+import java.util.LinkedHashMap;
+import java.util.List;
+import java.util.Map;
+import java.util.concurrent.ExecutionException;
+import java.util.concurrent.ExecutorService;
+import java.util.concurrent.Future;
+import java.util.concurrent.LinkedBlockingQueue;
+import java.util.concurrent.RejectedExecutionException;
+import java.util.concurrent.ThreadPoolExecutor;
+import java.util.concurrent.TimeUnit;
+import java.util.concurrent.TimeoutException;
+
+/**
+ * Probes the interpreter processes of one interpreter setting and reports whether they answer,
+ * within a deadline that holds for the whole request.
+ *
+ * Unlike the interpreter status snapshot, which is built from server memory alone, this does
+ * contact the interpreter. That is only acceptable because a probe happens when a user explicitly
+ * asks for one, against a single setting, and under a deadline: {@code isRunning()} costs a socket
+ * connect for docker and a kube-apiserver round trip for k8s, so the same call would be an accident
+ * if a listing made it on a timer.
+ *
+ * The deadline can only be imposed from the outside. {@code isRunning()} takes no timeout
+ * argument, the socket connect behind the docker implementation hardcodes one second, and the
+ * kubernetes client is created with the defaults of its library. So the probe runs on another
+ * thread and this class stops waiting for it, which bounds the response but not the probe:
+ * {@code Future#cancel(boolean)} interrupts, and neither a blocking socket connect nor an HTTP
+ * client call ends on an interrupt. A probe thread can therefore stay busy after its result was
+ * given up on, which is why the pool is small and its threads are daemons.
+ */
+public class InterpreterHealthChecker {
+
+ private static final Logger LOGGER = LoggerFactory.getLogger(InterpreterHealthChecker.class);
+
+ /**
+ * Budget for one health check request, shared by all groups of the setting rather than granted to
+ * each of them, so that an isolated setting cannot turn into a per group multiple of it.
+ */
+ @VisibleForTesting
+ static final long PROBE_TIMEOUT_IN_MILLIS = 3_000;
+
+ /**
+ * A probe that was given up on keeps its thread, so this bounds how many can pile up while an
+ * interpreter host is unreachable. Further probes wait for a thread instead of getting one, and
+ * are reported as timed out by the deadline if they never start, which is the honest answer.
+ */
+ private static final int MAX_CONCURRENT_PROBES = 4;
+
+ private final long probeTimeoutInMillis;
+ private final ExecutorService probeExecutor;
+
+ public InterpreterHealthChecker() {
+ this(PROBE_TIMEOUT_IN_MILLIS);
+ }
+
+ @VisibleForTesting
+ InterpreterHealthChecker(long probeTimeoutInMillis) {
+ this.probeTimeoutInMillis = probeTimeoutInMillis;
+ this.probeExecutor = newProbeExecutor();
+ }
+
+ private static ExecutorService newProbeExecutor() {
+ ThreadPoolExecutor executor = new ThreadPoolExecutor(
+ MAX_CONCURRENT_PROBES, MAX_CONCURRENT_PROBES, 60L, TimeUnit.SECONDS,
+ new LinkedBlockingQueue<>(),
+ runnable -> {
+ Thread thread = new Thread(runnable, "InterpreterHealthChecker-probe");
+ thread.setDaemon(true);
+ return thread;
+ });
+ executor.allowCoreThreadTimeOut(true);
+ return executor;
+ }
+
+ /**
+ * Probes every interpreter group of the given setting.
+ *
+ * A setting owns one group per interpreter process, so an isolated setting is probed once per
+ * user or note rather than once in total. The probes are submitted together and then collected
+ * against one deadline, so a group that does not answer costs the request its budget once instead
+ * of delaying the groups that would have answered right away.
+ */
+ public InterpreterHealthCheck check(InterpreterSetting interpreterSetting) {
+ List This is the one interpreter read path that does contact the interpreter, which is acceptable
+ * because a user asks for it explicitly, it names a single setting, and the server bounds the
+ * wait. It never starts an interpreter: a setting with nothing running is reported as such,
+ * since that is the normal state right after a restart.
+ *
+ * @param message HealthCheckInterpreterRequest, optional
+ */
+ @POST
+ @Path("setting/{settingId}/healthcheck")
+ @ZeppelinApi
+ public Response healthCheckSetting(String message, @PathParam("settingId") String settingId) {
+ InterpreterSetting setting = interpreterSettingManager.get(settingId);
+ if (setting == null) {
+ return new JsonResponse<>(Status.NOT_FOUND, "", settingId).build();
+ }
+
+ HealthCheckInterpreterRequest request =
+ GSON.fromJson(message, HealthCheckInterpreterRequest.class);
+ String noteId = request == null ? null : request.getNoteId();
+ if (null != noteId && !hasPermissionOnNote(noteId)) {
+ return new JsonResponse<>(Status.FORBIDDEN, "No privilege to health check interpreter")
+ .build();
+ }
+
+ LOGGER.info("Health check interpreterSetting {}, user={}", settingId,
+ authenticationService.getPrincipal());
+ return new JsonResponse<>(Status.OK, "", interpreterSettingManager.healthCheck(setting))
+ .build();
+ }
+
+ /**
+ * @return whether the current user may act on the interpreters of the given note, which running a
+ * paragraph of it already implies
+ */
+ private boolean hasPermissionOnNote(String noteId) {
+ Set The note is what a caller who is not an administrator is authorized against, the same way a
+ * restart from a note page is. A request without one is only served to whoever the deployment lets
+ * reach the interpreter endpoints at all.
+ */
+public class HealthCheckInterpreterRequest {
+
+ private final String noteId;
+
+ public HealthCheckInterpreterRequest(String noteId) {
+ this.noteId = noteId;
+ }
+
+ public String getNoteId() {
+ return noteId;
+ }
+}
diff --git a/zeppelin-server/src/test/java/org/apache/zeppelin/interpreter/InterpreterHealthCheckerTest.java b/zeppelin-server/src/test/java/org/apache/zeppelin/interpreter/InterpreterHealthCheckerTest.java
new file mode 100644
index 00000000000..8d4dc03c467
--- /dev/null
+++ b/zeppelin-server/src/test/java/org/apache/zeppelin/interpreter/InterpreterHealthCheckerTest.java
@@ -0,0 +1,215 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You 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
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * 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.
+ */
+
+package org.apache.zeppelin.interpreter;
+
+import org.apache.zeppelin.interpreter.InterpreterHealthCheck.GroupHealth;
+import org.apache.zeppelin.interpreter.InterpreterHealthCheck.Reason;
+import org.apache.zeppelin.interpreter.remote.RemoteInterpreterProcess;
+import org.junit.jupiter.api.AfterEach;
+import org.junit.jupiter.api.Test;
+
+import java.util.Arrays;
+import java.util.Collections;
+import java.util.List;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertFalse;
+import static org.junit.jupiter.api.Assertions.assertNull;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+import static org.mockito.ArgumentMatchers.any;
+import static org.mockito.ArgumentMatchers.anyString;
+import static org.mockito.Mockito.mock;
+import static org.mockito.Mockito.never;
+import static org.mockito.Mockito.verify;
+import static org.mockito.Mockito.when;
+
+/**
+ * Tests the on demand health check, above all that it stays within its deadline and that it never
+ * starts an interpreter to answer.
+ *
+ * The deadline is what makes probing acceptable here at all: {@code isRunning()} has no timeout
+ * of its own, so without it a single unreachable interpreter would hold the request for as long as
+ * the launcher underneath takes to give up.
+ */
+class InterpreterHealthCheckerTest {
+
+ /** Short enough to keep the tests quick, long enough that a local probe finishes well inside. */
+ private static final long PROBE_TIMEOUT_IN_MILLIS = 500;
+
+ private InterpreterHealthChecker healthChecker;
+
+ @AfterEach
+ void tearDown() {
+ if (healthChecker != null) {
+ healthChecker.stop();
+ }
+ }
+
+ /**
+ * A setting whose interpreters are not running is the normal state right after a restart, so it
+ * is reported as such - and answering must not be what starts them.
+ */
+ @Test
+ void reportsNotRunningWithoutStartingAnInterpreter() throws Exception {
+ ManagedInterpreterGroup interpreterGroup = interpreterGroup("group-1", null);
+ InterpreterSetting interpreterSetting =
+ interpreterSetting(Collections.singletonList(interpreterGroup));
+ healthChecker = new InterpreterHealthChecker(PROBE_TIMEOUT_IN_MILLIS);
+
+ InterpreterHealthCheck healthCheck = healthChecker.check(interpreterSetting);
+
+ assertEquals(1, healthCheck.getGroups().size());
+ GroupHealth groupHealth = healthCheck.getGroups().get(0);
+ assertEquals(Reason.NOT_RUNNING, groupHealth.getReason());
+ assertFalse(groupHealth.isHealthy(), "a group without a process cannot be healthy");
+ assertNull(groupHealth.getAlive(), "nothing was probed, so nothing is known about the process");
+ assertNull(groupHealth.getRunning());
+ verify(interpreterGroup, never()).getOrCreateInterpreterProcess(anyString(), any());
+ }
+
+ /** A setting without any group at all reports on the setting rather than on groups. */
+ @Test
+ void reportsNotRunningWhenThereIsNoGroup() {
+ InterpreterSetting interpreterSetting = interpreterSetting(Collections.emptyList());
+ healthChecker = new InterpreterHealthChecker(PROBE_TIMEOUT_IN_MILLIS);
+
+ InterpreterHealthCheck healthCheck = healthChecker.check(interpreterSetting);
+
+ assertEquals(Reason.NOT_RUNNING, healthCheck.getReason());
+ assertTrue(healthCheck.getGroups().isEmpty());
+ }
+
+ /** A reachable interpreter is healthy, and saying so takes a single remote call. */
+ @Test
+ void reportsHealthyOnASingleRemoteCall() {
+ RemoteInterpreterProcess process = mock(RemoteInterpreterProcess.class);
+ when(process.isRunning()).thenReturn(true);
+ InterpreterSetting interpreterSetting = interpreterSetting(
+ Collections.singletonList(interpreterGroup("group-1", process)));
+ healthChecker = new InterpreterHealthChecker(PROBE_TIMEOUT_IN_MILLIS);
+
+ GroupHealth groupHealth = healthChecker.check(interpreterSetting).getGroups().get(0);
+
+ assertEquals(Reason.OK, groupHealth.getReason());
+ assertTrue(groupHealth.isHealthy());
+ assertEquals(Boolean.TRUE, groupHealth.getAlive());
+ // Asking whether the process exists would be another remote call for no added answer.
+ verify(process, never()).isAlive();
+ }
+
+ /**
+ * The point of the whole exercise: a probe that does not come back must not hold the request. The
+ * interpreter here never answers, which is what an unreachable host looks like.
+ */
+ @Test
+ void stopsWaitingForAProbeThatDoesNotComeBack() {
+ RemoteInterpreterProcess process = mock(RemoteInterpreterProcess.class);
+ when(process.isRunning()).thenAnswer(invocation -> {
+ Thread.sleep(60_000);
+ return true;
+ });
+ InterpreterSetting interpreterSetting = interpreterSetting(
+ Collections.singletonList(interpreterGroup("group-1", process)));
+ healthChecker = new InterpreterHealthChecker(PROBE_TIMEOUT_IN_MILLIS);
+
+ long startTimeInMillis = System.currentTimeMillis();
+ GroupHealth groupHealth = healthChecker.check(interpreterSetting).getGroups().get(0);
+ long tookInMillis = System.currentTimeMillis() - startTimeInMillis;
+
+ assertEquals(Reason.PROBE_TIMEOUT, groupHealth.getReason());
+ assertNull(groupHealth.getAlive(), "the probe never answered, so nothing is known");
+ assertTrue(tookInMillis < PROBE_TIMEOUT_IN_MILLIS * 10,
+ "the request should end on the deadline rather than with the probe, but took "
+ + tookInMillis + "ms");
+ }
+
+ /**
+ * The deadline belongs to the request, so the groups are probed together: one interpreter that
+ * does not answer must not cost the others their result.
+ */
+ @Test
+ void oneUnreachableGroupDoesNotHideTheOthers() {
+ RemoteInterpreterProcess reachable = mock(RemoteInterpreterProcess.class);
+ when(reachable.isRunning()).thenReturn(true);
+ RemoteInterpreterProcess unreachable = mock(RemoteInterpreterProcess.class);
+ when(unreachable.isRunning()).thenAnswer(invocation -> {
+ Thread.sleep(60_000);
+ return true;
+ });
+ InterpreterSetting interpreterSetting = interpreterSetting(Arrays.asList(
+ interpreterGroup("group-unreachable", unreachable),
+ interpreterGroup("group-reachable", reachable)));
+ healthChecker = new InterpreterHealthChecker(PROBE_TIMEOUT_IN_MILLIS);
+
+ List