Spark向Elasticsearch批量導(dǎo)入數(shù)據(jù)，出現(xiàn)重復(fù)記錄問題的定位

看了下es-hadoop插件的源碼：

發(fā)現(xiàn)ES導(dǎo)入數(shù)據(jù)重試情況的發(fā)生，除了在es.batch.write.retry.policy參數(shù)默認開啟且es-hadoop插件向ES集群發(fā)送bulk寫入請求接受到503響應(yīng)碼會重試3次之外。

本身執(zhí)行http請求時，也會存在重試（hadoop/rest/NetworkClient.java）：

   public Response execute(Request request) {
        Response response = null;
        boolean newNode;
        do {
            SimpleRequest routedRequest = new SimpleRequest(request.method(), null, request.path(), request.params(), request.body());

            newNode = false;
            try {
                response = currentTransport.execute(routedRequest);
                ByteSequence body = routedRequest.body();
                if (body != null) {
                    stats.bytesSent += body.length();
                }
            } catch (Exception ex) {
                // configuration error - including SSL/PKI - bail out
                if (ex instanceof EsHadoopIllegalStateException) {
                    throw (EsHadoopException) ex;
                }
                // issues with the SSL handshake, bail out instead of retry, for security reasons
                if (ex instanceof javax.net.ssl.SSLException) {
                    throw new EsHadoopTransportException(ex);
                }
                // check for fatal, non-recoverable network exceptions
                if (ex instanceof BindException) {
                    throw new EsHadoopTransportException(ex);
                }

                if (log.isTraceEnabled()) {
                    log.trace(
                            String.format(
                                    "Caught exception while performing request [%s][%s] - falling back to the next node in line...",
                                    currentNode, request.path()), ex);
                }

                String failed = currentNode;

                failedNodes.put(failed, ex);

                newNode = selectNextNode();

                log.error(String.format("Node [%s] failed (%s); "
                        + (newNode ? "selected next node [" + currentNode + "]" : "no other nodes left - aborting..."),
                        failed, ex.getMessage()));

                if (!newNode) {
                    throw new EsHadoopNoNodesLeftException(failedNodes);
                }
            }
        } while (newNode);

        return response;
    }

當請求出現(xiàn)超時的情況時，es-hadoop插件會再請求一個ES節(jié)點發(fā)送寫入請求。即導(dǎo)入插件認為當前插入節(jié)點超時了（默認是一分鐘）就視為該節(jié)點不可用，就換下一個節(jié)點，其實是ES在一分鐘內(nèi)沒有處理完插入任務(wù)。

將超時時間es.http.timeout參數(shù)調(diào)大之后，給ES留下充足的入庫時間，就不會再發(fā)生這個問題了。