Question

我在AWS上设置了一个Cassandra集群。我想要的是增加更多节点（如所宣传的），增加I / O吞吐量（每秒读/写次数）。但是，我恰恰相反。随着新节点的添加，性能会降低。

您是否知道阻止其缩放的任何典型问题？

以下是一些细节：

我正在为列族添加一个文本文件（15MB）。每一行都是一条记录。有150000条记录。当有1个节点时，写入大约需要90秒。但是当有2个节点时，需要120秒。我可以看到数据传播到2个节点。但是，吞吐量没有增加。

源代码如下：

public class WordGenCAS {
static final String KEYSPACE = "text_ks";
static final String COLUMN_FAMILY = "text_table";
static final String COLUMN_NAME = "text_col";

public static void main(String[] args) throws Exception {
    if (args.length < 2) {
        System.out.println("Usage: WordGenCAS <input file> <host1,host2,...>");
        System.exit(-1);
    }

    String[] contactPts = args[1].split(",");

    Cluster cluster = Cluster.builder()
            .addContactPoints(contactPts)
            .build();
    Session session = cluster.connect(KEYSPACE);

    InputStream fis = new FileInputStream(args[0]);
    InputStreamReader in = new InputStreamReader(fis, "UTF-8");
    BufferedReader br = new BufferedReader(in);

    String line;
    int lineCount = 0;
    while ( (line = br.readLine()) != null) {
        line = line.replaceAll("'", " ");
        line = line.trim();
        if (line.isEmpty())
            continue;
        System.out.println("[" + line + "]");
        String cqlStatement2 = String.format("insert into %s (id, %s) values (%d, '%s');",
                COLUMN_FAMILY,
                COLUMN_NAME,
                lineCount,
                line);
        session.execute(cqlStatement2);
        lineCount++;
    }

    System.out.println("Total lines written: " + lineCount);
}

}

数据库架构如下：

CREATE KEYSPACE text_ks WITH REPLICATION = { 'class' : 'SimpleStrategy', 'replication_factor' : 2 };

USE text_ks;

CREATE TABLE text_table (
    id int,
    text_col text,
    primary key (id)
) WITH COMPACT STORAGE;

谢谢！

Answer 1

即使这是一篇旧帖子，我认为值得为这些（常见）问题发布解决方案。

正如您已经发现的那样，使用 serial 过程加载数据的速度很慢。你被建议的是正确的做法。

但是，在不施加某种背压的情况下发出大量查询可能会导致问题，并且由于服务器（以及驱动程序在某种程度上）过度过载，您将丢失数据。

此解决方案将使用异步调用加载数据，并尝试在客户端上施加一些背压以避免数据丢失。

public class WordGenCAS {
    static final String KEYSPACE = "text_ks";
    static final String COLUMN_FAMILY = "text_table";
    static final String COLUMN_NAME = "text_col";

    public static void main(String[] args) throws Exception {
        if (args.length < 2) {
            System.out.println("Usage: WordGenCAS <input file> <host1,host2,...>");
            System.exit(-1);
        }

        String[] contactPts = args[1].split(",");

        Cluster cluster = Cluster.builder()
                .addContactPoints(contactPts)
                .build();
        Session session = cluster.connect(KEYSPACE);

        InputStream fis = new FileInputStream(args[0]);
        InputStreamReader in = new InputStreamReader(fis, "UTF-8");
        BufferedReader br = new BufferedReader(in);

        String line;
        int lineCount = 0;

        // This is the futures list of our queries
        List<Future<ResultSet>> futures = new ArrayList<>();

        // Loop
        while ( (line = br.readLine()) != null) {
            line = line.replaceAll("'", " ");
            line = line.trim();
            if (line.isEmpty())
                continue;
            System.out.println("[" + line + "]");
            String cqlStatement2 = String.format("insert into %s (id, %s) values (%d, '%s');",
                    COLUMN_FAMILY,
                    COLUMN_NAME,
                    lineCount,
                    line);
            lineCount++;

            // Add the "future" returned by async method the to the list
            futures.add(session.executeAsync(cqlStatement2));

            // Apply some backpressure if we issued more than X query.
            // Change X to another value suitable for your cluster
            while (futures.size() > 1000) {
                Future<ResultSet> future = futures.remove(0);
                try {
                    future.get();
                } catch (Exception e) {
                    e.printStackTrace();
                }
            }
        }

        System.out.println("Total lines written: " + lineCount);
        System.out.println("Waiting for writes to complete...");

        // Wait until all writes are done.
        while (futures.size() > 0) {
            Future<ResultSet> future = futures.remove(0);
            try {
                future.get();
            } catch (Exception e) {
                e.printStackTrace();
            }
        }

        System.out.println("Done!");
    }
}

Cassandra减速了更多的节点

1 个答案: