lento ¹

Streaming Node.js client for Presto, the "Distributed SQL Query Engine for Big Data".

Table of Contents

Click to expand

• Features
• Example
• API
• Install
• License

Features

• Stream rows or pages of rows
• Cancelation through stream.destroy()
• Set session properties
• Set time limit (enforced by Presto)
• Get rows as objects or arrays
• Supports both callbacks & promises
• Uses Presto HTTP protocol v1
• Keep-alive HTTP connections
• Supports Gzip and Deflate content encoding
• Retries HTTP 503 and other failures.

Example

Convert a Presto table to CSV:

const lento = require('lento')
const csvWriter = require('csv-write-stream')
const stdout = require('stdout-stream')
const pipeline = require('readable-stream').pipeline

const client = lento({
  hostname: 'example',
  port: 8080,
  catalog: 'hive',
  schema: 'test',
  user: 'test'
})

const source = client.createRowStream('SELECT * FROM events')
const transform = csvWriter()

pipeline(source, transform, stdout, (err) => {
  if (err) throw err
})

If the destination streams close or error, the source stream is destroyed (courtesy of pipeline) and the Presto query canceled.

API

• lento([options])
• createPageStream(sql[, options])
  • Cancelation
  • Events
• createRowStream(sql[, options])
• query(sql[, options], callback)
• setTimeout(duration[, options], callback)
• resetTimeout([options, ]callback)
• set(key, value[, options], callback)
• reset(key[, options], callback)
• session([options, ]callback)
• Errors
• Builtin Retry
• Debug

lento([options])

Options:

• hostname: string, default is localhost
• port: number, default 8080
• protocol: string, one of http: (default) or https:
• user: string, default none. Sent as X-Presto-User header.
• timezone: string, for example UTC, default none. Sent as X-Presto-Time-Zone header.
• headers: object containing custom headers to set on every request such as Authorization (case-insensitive). Headers specified here take precedence over other options that set headers.
• parametricDatetime: boolean, default false. Opt-in to datetime types with variable precision, for example timestamp(6). When not set, datetime types are returned with a precision of 3.

You can specify a catalog and schema to avoid writing fully-qualified table names in queries:

• catalog: string, for example hive, default none. Sent as X-Presto-Catalog header.
• schema: string, for example logs, default none. Sent as X-Presto-Schema header.

Control delays and retries:

• pollInterval: number (milliseconds) or string with unit (e.g. 1s, 500ms). How long to wait for server-side state changes, before sending another HTTP request. Default is 1 second.
• maxRetries: number of retries if Presto responds with HTTP 503 or other failures. Default is 10.
• socketTimeout: number (milliseconds) or string with unit. When to timeout after inactivity on the socket. Default is 2 minutes.

createPageStream(sql[, options])

Execute a query. Takes sql as a string or Buffer and returns a readable stream that yields pages of rows.

const through2 = require('through2')

client
  .createPageStream('SELECT * FROM events')
  .pipe(through2.obj((page, enc, next) => {
    for (let row of page) {
      // ..
    }

    // Process next page
    next()
  }))

Options:

• pageSize: number, default 500
• highWaterMark: number, default 0
• rowFormat: string, one of object (default) or array
• deserialize: boolean, default true
• headers: custom request headers. Merged with headers that were set in the constructor, if any.

The pageSize specifies the maximum number of rows per page. Presto may return less per page. If Presto returns more rows than pageSize, the surplus is buffered and the stream will not make another HTTP request to Presto until fully drained. Note that if the (remainder of) rows fit in Presto's buffers, Presto will not block (until another HTTP request is made) but instead go into the FINISHED state after which you have 15 minutes (by default) to fetch the remaining results. If pageSize is <= 0 the stream emits pages as returned by Presto, without slicing them up.

The highWaterMark specifies how many pages to fetch preemptively. The maximum numbers of rows held in memory is approximately (highWaterMark || 1) * pageSize, plus any surplus from the last HTTP request. Because Presto can return thousands of rows per request, the default highWaterMark is 0 so that we don't preemptively fetch and only hold the number of rows contained in the last HTTP request.

If you care more about throughput, you can opt to increase highWaterMark. Additionally you can increase pageSize if processing time is minimal or if you don't mind blocking the rest of your app while processing a page.

For tuning the Presto side of things, use set().

Cancelation

Destroying the stream will cancel the query with a DELETE request to Presto, unless no requests were made yet. If the initial request is in flight the stream will wait for a response, which contains the query id that can then be cancelled. If cancelation fails the stream will emit an error (open an issue if you think it shouldn't). Regardless of success, the streams emits close as the last event.

Events

Besides the usual Node.js stream events, the stream emits:

• id: emitted with query id once known
• info: emitted with fully qualified info URL
• columns: emitted with an array of column metadata as returned by Presto
• stats: emitted for each HTTP response, with raw data
• state_change: emitted with a string state (e.g. RUNNING, FINISHED) when Presto state changes. Should not be relied upon, only meant for debugging.
• raw_page_size: emitted with a count of rows (e.g. 21829) for each HTTP response that has rows. For debugging.

Note The id, info and columns events may be emitted more than once due to retries. Subsequent stats and state_change events pertain to that retried query as well.

createRowStream(sql[, options])

Execute a query. Takes sql as a string or Buffer and returns a readable stream that yields rows. Options:

• highWaterMark: number, default 16
• rowFormat: string, one of object (default) or array
• deserialize: boolean, default true
• headers: custom request headers. Merged with headers that were set in the constructor, if any.

query(sql[, options][, callback])

Same as above but non-streaming, meant for simple queries. The callback function will receive an error if any and an array of rows. If no callback is provided, a promise is returned.

client.query('DESCRIBE events', (err, rows) => {
  // ..
})

With async/await:

const rows = await client.query('DESCRIBE events')

setTimeout(duration[, options][, callback])

Set query_max_run_time for subsequent queries. If those take longer than duration, Presto will return an error with code EXCEEDED_TIME_LIMIT (see errors). The duration can be a number (in milliseconds) or a string parsed by Presto with the format <value><unit> - for example 5d or 100ms. Options are passed to query() via set(), as this method is a shortcut for:

client.set('query_max_run_time', duration[, options], callback)

If no callback is provided, a promise is returned.

resetTimeout([options][, callback])

Reset query_max_run_time to Presto's default. Options are passed to query() via reset(). If no callback is provided, a promise is returned.

set(key, value[, options][, callback])

Set a session property. Executes SET SESSION .. to prevalidate input, then sets X-Presto-Session header on subsequent queries. Value can be a boolean, number or string. Options are passed to query(). If no callback is provided, a promise is returned.

client.set('redistribute_writes', false, (err) => {
  if (err) return console.error('failed to set', err)

  // Subsequent queries now use redistribute_writes=false
})

reset(key[, options][, callback])

Reset a session property to its default value. Options are passed to query(). If no callback is provided, a promise is returned.

client.reset('redistribute_writes', (err) => {
  if (err) return console.error('failed to reset', err)

  // Subsequent queries now use the default value of redistribute_writes
})

session([options, ][callback])

Converts the result of SHOW SESSION into a tree, coerces boolean and integer values to JSON types. Options are passed to query(). Callback signature is (err, session). If no callback is provided, a promise is returned.

Partial example of a session:

{
  "execution_policy": {
    "key": "execution_policy",
    "value": "all-at-once",
    "default": "all-at-once",
    "type": "varchar",
    "description": "Policy used for scheduling query tasks"
  },
  "hash_partition_count": {
    "key": "hash_partition_count",
    "value": 100,
    "default": 100,
    "type": "integer",
    "description": "Number of partitions for distributed joins and aggregations"
  },
  "hive": {
    "bucket_execution_enabled": {
      "key": "hive.bucket_execution_enabled",
      "value": true,
      "default": true,
      "type": "boolean",
      "description": "Enable bucket-aware execution: only use a single worker per bucket"
    }
  }
}

See Presto Properties for a detailed description of (some of) the properties.

Errors

Errors are enriched with a code and type (string) and optionally info. For example:

client.setTimeout('1ms', (err) => {
  if (err) throw err

  client.query('SELECT * FROM big_table', (err) => {
    console.error(err.message) // 'EXCEEDED_TIME_LIMIT: Query exceeded maximum time limit of 1.00ms'
    console.error(err.code) // 'EXCEEDED_TIME_LIMIT'
    console.error(err.type) // 'INSUFFICIENT_RESOURCES'
  })
})

Builtin Retry

If Presto responds to an HTTP request with 503 or if the TCP connection is refused, lento retries the request with an exponential delay between 1 and 10 seconds.

In addition, a query (consisting of one or more HTTP requests) will be retried if Presto returns an error like SERVER_STARTING_UP or HIVE_METASTORE_ERROR, but only if no data was received yet, to avoid emitting duplicates.

I wish retries could be handled at a higher level, but as it stands, lento is both a low-level HTTP client and a streaming client, so retries have to be handled here. This may change in the future.

Debug

Enable debug output with DEBUG=lento. Mostly logs HTTP requests and retries, no usernames, SQL or other potentially sensitive data. Beware, it can log hundreds of lines per query.

Install

With npm do:

npm install lento

License

MIT © 2018-present Vincent Weevers

---

¹ Because streams are about slowing down! ↩