CLI

Command line flags

The materialized binary supports the following command line flags:

If a command line flag takes an argument, you can alternatively set that flag via an environment variable named after the flag. If both the environment variable and command line flag are specified, the command line flag takes precedence.

The process for converting a flag name to an environment variable name is as follows:

1. Convert all characters to uppercase
2. Replace all hyphens with underscores
3. add an MZ_ prefix.

For example, the --data-directory command line flag corresponds to the MZ_DATA_DIRECTORY environment variable.

Note that command line flags that do not take arguments, like --experimental and --disable-telemetry, do not yet have corresponding environment variables.

Data directory

Upon startup materialized creates a directory where it persists metadata. By default, this directory is called mzdata and is situated in the current working directory of the materialized process. Currently, only metadata is persisted in mzdata. You can specify a different directory using the --data-directory flag. Upon start, materialized checks for an existing data directory, and will reinstall source and view definitions from it if one is found.

Worker threads

A materialized instance runs a specified number of timely dataflow worker threads. Worker threads can only be specified at startup by setting the --workers flag, and cannot be changed without shutting down materialized and restarting. If --workers is not set, materialized will default to using half of the machine’s physical cores as the thread count. In the future, dynamically changing the number of worker threads will be possible over distributed clusters, see #2449.

Changed in v0.4.0: Rename the --threads flag to --workers.

Changed in v0.5.1: When unspecified, default to using half of the machine’s physical cores.

How many worker threads should you run?

Adding worker threads allows Materialize to handle more throughput. Reducing worker threads consumes fewer resources, and reduces tail latencies.

In general, you should use the fewest number of worker threads that can handle your peak throughputs. This is also the most resource efficient.

You should never run Materialize in a configuration greater than n-1 workers, where n is the number of physical cores. Note that major cloud providers list the number of hyperthreaded cores (or virtual CPUs). Divide this number by two to get the number of physical cores available. The reasoning is simple: Timely Dataflow is very computationally efficient and typically uses all available computational resources. Under high throughput, you should see each worker pinning a core at 100% CPU, with no headroom for hyperthreading. One additional core is required for metadata management and coordination. Timely workers that have to fight for physical resources will only block each other.

Example: an r5d.4xlarge instance has 16 VCPUs, or 8 physical cores. The recommended worker setting on this VM is 7.

Listen address

By default, materialized binds to 0.0.0.0:6875. This means that Materialize will accept any incoming SQL connection to port 6875 from anywhere. It is the responsibility of the network firewall to limit incoming connections. If you wish to configure materialized to only listen to, e.g. localhost connections, you can set --listen-addr to localhost:6875. You can also use this to change the port that Materialize listens on from the default 6875.

Compaction window

The --logical-compaction-window option specifies the duration of time for which Materialize is required to maintain full historical detail in its arrangements. Note that compaction happens lazily, so Materialize may retain more historical detail than requested, but it will never retain less.

The value of the option is any valid SQL interval string, like 10ms (10 milliseconds) or 1min 30s (1 minute, 30 seconds). The special value off disables logical compaction and corresponds to an unboundedly large duration.

The logical compaction window ends at the current time and extends backwards in time for the configured duration. The default window is 1 millisecond.

See the Deployment section for guidance on tuning the compaction window.

Logging

Log file

The --log-file option specifies the path to a file in which Materialize will write its log messages. The value stderr is treated specially and specifies the standard error stream.

If the option is unspecified, Materialize writes log messages to the materialized.log file in the data directory and additionally forwards any log messages at the WARN or ERROR levels to the standard error stream. Forwarding does not occur if you explicitly specify a log file.

Log filter

New in v0.7.2.

The --log-filter option specifies which log messages Materialize will emit. Its value is a comma-separated list of filter directives. Each filter directive has the following format:

[module::path=]level

A filter directive registers interest in log messages from the specified module that are at least as severe as the specified level. If a directive omits the module, then it implicitly applies to all modules. When directives conflict, the last directive wins. Materialize will only emit log messages that match at least one filter directive.

Specifying module paths in filter directives requires familiarity with Materialize’s codebase and is intended for advanced users.

The valid levels for a log message are documented in the logging section of the monitoring documentation and are not case sensitive. The special level off may be used in a directive to suppress all log messages, even those at the error level.

As an example, the following filter specifies the TRACE level for the pgwire module, which handles SQL network connections, and the INFO level for all other modules.

pgwire=trace,info

Introspection sources

Changed in v0.7.1: In prior versions of Materialize, this option was undocumented but available under the name --logging-granularity.

Materialize maintains several built-in sources and views in mz_catalog that describe the internal state of the dataflow execution layer, like mz_scheduling_elapsed.

The --introspection-frequency option determines the frequency at which the base sources are updated. The default frequency is 1s. To disable introspection entirely, use the special value off.

Higher frequencies provide more up-to-date introspection but increase load on the system. Lower frequencies increase staleness in exchange for decreased load. The default frequency is a good choice for most deployments.

Prometheus metrics

Changed in v0.9.1: In prior versions of Materialize, the metrics scraping interval was linked to the introspection interval.

The --metrics-scraping-interval option determines the interval at which the prometheus metrics are collected to update the mz_metrics table. The default interval is 30s. To disable prometheus metrics collection entirely, use the special value off.

Lower intervals provide more up-to-date metrics but increase load on the system. Higher intervals increase staleness in exchange for decreased load. The default interval is a good choice for most deployments.

Changed in v0.7.3: Materialize imports its own Prometheus metrics into the systems tables mz_metrics (counters and gauge readings), mz_metric_histograms (histogram distributions) and mz_metrics_meta (type information and help for each metric). These readings are imported once per --metrics-scraping-interval period, and are retained for the duration given with --retain-prometheus-metrics (defaulting to 5 minutes). Higher retention periods lead to greater memory usage.

TLS encryption

Materialize can use Transport Layer Security (TLS) to:

• Encrypt traffic between SQL and HTTP clients and the materialized server
• Authenticate SQL and HTTP clients

New in v0.7.1: The --tls-mode and --tls-ca options.

Configuration

Whether Materialize requires TLS encryption or authentication is determined by the value of the --tls-mode option:

In all TLS modes but disable, you will need to supply two files, one containing a TLS certificate and one containing the corresponding private key. Point materialized at these files using the --tls-cert and --tls-key options, respectively.

If the TLS mode is verify-ca or verify-full, you will additionally need to supply the path to a TLS certificate authority (CA) via the --tls-ca flag. Client certificates will be verified using this CA.

The following example demonstrates how to configure a server in verify-full mode:

$ materialized -w1 --tls-cert=server.crt --tls-key=server.key --tls-ca=root.crt

Materialize statically links against a vendored copy of OpenSSL. It does not use any SSL library that may be provided by your system. To see the version of OpenSSL used by a particular materialized binary, inquire with the -vv flag:

$ materialize -vv

materialized v0.2.3-dev (c62c988e8167875b92122719eee5709cf81cdac4)
OpenSSL 1.1.1g  21 Apr 2020
librdkafka v1.4.2

Materialize configures OpenSSL according to Mozilla’s Intermediate compatibility level, which requires TLS v1.2+ and recent cipher suites. Using weaker cipher suites or older TLS protocol versions is not supported.

Generating TLS certificates

You can generate a self-signed certificate for development use with the openssl command-line tool:

$ openssl req -new -x509 -days 365 -nodes -text \
    -out server.crt -keyout server.key -subj "/CN=<SERVER-HOSTNAME>"

Production deployments typically should not use self-signed certificates. Acquire a certificate from a proper certificate authority (CA) instead.

Experimental mode

New in v0.4.0.

Materialize offers access to experimental features through the --experimental flag. Unlike most features in Materialize, experimental features’ syntax and/or semantics can shift at any time, and there is no guarantee that future versions of Materialize will be interoperable with the experimental features.

Using experimental mode means that you are likely to lose access to all of your sources and views within Materialize and will have to recreate them and re-ingest all of your data.

Because of this volatility:

• You can only initialize new servers in experimental mode.
• Servers started in experimental mode must always be started in experimental mode.

We recommend only using experimental mode to explore Materialize, i.e. absolutely never in production. If your explorations yield interesting results or things you’d like to see changed, let us know on GitHub.

Disabling experimental mode

You cannot disable experimental mode for a server. You can, however, extract your view and source definitions (SHOW CREATE VIEW, SHOW CREATE SOURCE, etc.), and then create a new server with those items.

Telemetry

Materialize periodically communicates with telemetry.materialize.com to report usage data and check for new versions. You can opt out of this communication with the --disable-telemetry flag.

We record the following data:

• Public IP of the host running Materialize
• Cluster ID, a unique ID which is persistent across Materialize restarts
• Session ID, a unique ID which is reset on each Materialize restart
• Materialize version
• Number of worker threads
• Uptime
• Count of sinks, sources, and views by type

We use this data to guide our product roadmap. Unless you are using Materialize Cloud, we do not and cannot correlate this data to your identity.

Dataflow tuning

There are several command-line options that tune various parameters for Materialize’s underlying dataflow engine:

• --differential-idle-merge-effort controls how aggressively Materialize will perform compaction when idle.
• --timely-progress-mode sets Timely Dataflow’s progress tracking mode.

Using these parameters correctly requires substantial knowledge about how the underlying Timely and Differential Dataflow engines work. Typically you should only set these parameters in consultation with Materialize engineers.

Disable user indexes

New in v0.9.2.

If you cannot boot a Materialize server because it runs out of memory, you can use the --disable-user-indexes to prevent Materialize from creating any indexes on user-created objects. For example, if you add a view that contains a cross join that causes your server to immediately run out of memory on boot, you can use --disable-user-indexes to boot the server and then drop the offending view.

In this mode users…

• Can access objects within the system catalog to help determine which indexes are causing the crash.
• Can only SELECT from user-created objects that do not rely on user-created indexes. In essence, this means users can still SELECT from user-created…
  • Tables, but they will never return any data.
  • Views that contain only references to constant values or depend entirely on system tables’ indexes.
• Cannot INSERT data into tables.
• Can create new objects, but any created indexes are disabled.

After troubleshooting any issues, you can enable individual indexes or reboot Materialize without --disable-user-indexes to enable all indexes at once.

For assistance with this mode, see:

• Architecture over: Indexes
• System catalog
• SHOW INDEX
• DROP VIEW
• DROP INDEX

Special environment variables

Materialize respects several environment variables that have conventional meanings in Unix systems.

HTTP proxies

The http_proxy, https_proxy, all_proxy, and no_proxy environment variables specify a proxy to use for outgoing HTTP and HTTPS traffic. There is no precise specification of how these variables behave, but Materialize’s behavior generally matches the behavior of other HTTP clients.

For precise details of Materialize’s behavior, consult the documentation of the mz_http_proxy crate.