Input Framework

Zeek features a flexible input framework that allows users to import arbitrary data into Zeek. Data is either read into Zeek tables or directly converted to events for scripts to handle as they see fit. A modular reader architecture allows reading from files, databases, or other data sources.

This chapter gives an overview of how to use the input framework, with examples. For more complex scenarios take a look at the test cases in testing/btest/scripts/base/frameworks/input/ in the Zeek distribution.

Note

The input framework has no awareness of Zeek’s cluster architecture. Zeek supports all of the mechanisms covered below on any cluster node. The config and intelligence frameworks both leverage the input framework, adding logic that applies the input framework on the manager node, distributing ingested information across the cluster via events.

Reading Data into Tables

Probably the most interesting use-case of the input framework is to read data into a Zeek table. By default, the input framework reads the data in the same format as it is written by Zeek’s logging framework: a tab-separated ASCII file.

We will show the ways to read files into Zeek with a simple example. For this example we assume that we want to import data from a denylist that contains server IP addresses as well as the timestamp and the reason for the block.

An example input file could look like this (note that all fields must be tab-separated):

#fields ip timestamp reason
192.168.17.1 1333252748 Malware host
192.168.27.2 1330235733 Botnet server
192.168.250.3 1333145108 Virus detected

To read a file into a Zeek table, two record types have to be defined. One contains the types and names of the columns that should constitute the table keys, and the second contains the types and names of the columns that should constitute the table values.

In our case, we want to be able to look up IPs. Hence, our key record only contains the server IP. All other elements should be stored as the table content.

type Idx: record {
    ip: addr;
};

type Val: record {
    timestamp: time;
    reason: string;
};

Note that the names of the fields in the record definitions must correspond to the column names listed in the #fields line of the input file, in this case ip, timestamp, and reason. Also note that the ordering of the columns does not matter, because each column is identified by name.

The input file is read into the table with a call of the Input::add_table function:

global denylist: table[addr] of Val = table();

event zeek_init() {
    Input::add_table([$source="denylist.file", $name="denylist",
                      $idx=Idx, $val=Val, $destination=denylist]);
    Input::remove("denylist");
}

With these three lines we first create an empty table that should receive the denylist data and then instruct the input framework to open an input stream named “denylist” to read the data into the table. The third line removes the input stream again, because we do not need it any more after the data has been read.

Note that while the key and content records may use &optional fields, omitting columns (usually via the “-” character) requires care. Since the key record’s columns expand into a list of values for indexing into the receiving table (note how in the above example denylist is indexed via a plain addr) and all of those values must be present for indexing, you cannot in practice omit these values. For content records, omitting is meaningful, but only permitted for columns with the &optional attribute. The framework skips offending input lines with a warning.

Note

Prior to version 4.1 Zeek accepted such inputs, unsafely. When transitioning from such versions to Zeek 4.1 or newer, users with omitted fields in their input data may observe discrepancies in the loaded data sets.

Asynchronous processing

Since some data files might be rather large, the input framework works asynchronously. A new thread is created for each new input stream. This thread opens the input data file, converts the data into an internal format and sends it back to the main Zeek thread. Because of this, the data is not immediately accessible. Depending on the size of the data source it might take from a few milliseconds up to a few seconds until all data is present in the table. Please note that this means that when Zeek is running without an input source or on very short captured files, it might terminate before the data is present in the table (because Zeek already handled all packets before the import thread finished).

Subsequent calls to an input source are queued until the previous action has been completed. Because of this it is, for example, possible to call Input::add_table and Input::remove in two subsequent lines: the remove action will remain queued until the first read has been completed.

Once the input framework finishes reading from a data source, it fires the Input::end_of_data event. Once this event has been received all data from the input file is available in the table.

event Input::end_of_data(name: string, source: string) {
    # now all data is in the table
    print denylist;
}

The table can be used while the data is still being read — it just might not contain all lines from the input file before the event has fired. After the table has been populated it can be used like any other Zeek table and denylist entries can easily be tested:

if ( 192.168.18.12 in denylist )
    # take action

Sets instead of tables

For some use cases the key/value notion that drives tabular data does not apply, for example when the main purpose of the data is to test for membership in a set. The input framework supports this approach by using sets as the destination data type, and omitting $val in Input::add_table:

type Idx: record {
    ip: addr;
};

global denylist: set[addr] = set();

event zeek_init() {
    Input::add_table([$source="denylist.file", $name="denylist",
                     $idx=Idx, $destination=denylist]);
    Input::remove("denylist");
}

Re-reading and streaming data

For some data sources (such as many denylists), the input data changes continually. The input framework supports additional techniques to manage such ever-changing input.

The first, very basic method is an explicit refresh of an input stream. When an input stream is open (meaning it has not yet been removed by a call to Input::remove), the function Input::force_update can be called. This will trigger a complete refresh of the table: any changed elements from the file will be updated, new ones added, and any elements no longer in the input data get removed. After the update is finished the Input::end_of_data event will be raised.

In our example the call would look as follows:

Input::force_update("denylist");

Alternatively, the input framework can automatically refresh the table contents when it detects a change to the input file. To use this feature you need to specify a non-default read mode by setting the mode option of the Input::add_table call. Valid values are Input::MANUAL (the default), Input::REREAD, and Input::STREAM. For example, setting the value of the mode option in the previous example would look like this:

Input::add_table([$source="denylist.file", $name="denylist",
                  $idx=Idx, $val=Val, $destination=denylist,
                  $mode=Input::REREAD]);

When using the reread mode (i.e., $mode=Input::REREAD), Zeek continually checks if the input file has been changed. If the file has been changed, it is re-read and the data in the Zeek table is updated to reflect the current state. Each time a change has been detected and all the new data has been read into the table, the Input::end_of_data event is raised.

When using the streaming mode (i.e., $mode=Input::STREAM), Zeek assumes that the input is an append-only file to which new data is continually appended. Zeek continually checks for new data at the end of the file and will add the new data to the table. If newer lines in the file have the same table index as previous lines, they will overwrite the values in the output table. Because of the nature of streaming reads (data is continually added to the table), the Input::end_of_data event is never raised when using streaming reads.

Tip

Change detection happens via periodic “heartbeat” events, defaulting to a frequency of once per second as defined by the global Threading::heartbeat_interval constant. The reader considers the input file changed when the file’s inode or modification time has changed since the last check.

Receiving change events

When re-reading files, it might be interesting to know exactly which lines in the source files have changed. For this reason, the input framework can raise an event each time when a data item is added to, removed from, or changed in a table.

The event definition looks like this (note that you can change the name of this event in your own Zeek script):

event entry(description: Input::TableDescription, tpe: Input::Event,
            left: Idx, right: Val) {
    # do something here...
    print fmt("%s = %s", left, right);
}

The event must be specified in $ev in the Input::add_table call:

Input::add_table([$source="denylist.file", $name="denylist",
                  $idx=Idx, $val=Val, $destination=denylist,
                  $mode=Input::REREAD, $ev=entry]);

The description argument of the event contains the arguments that were originally supplied to the Input::add_table call. Hence, the name of the stream can, for example, be accessed with description$name. The tpe argument of the event is an enum containing the type of the change that occurred.

If a line that was not previously present in the table has been added, then the value of tpe will be Input::EVENT_NEW. In this case left contains the index of the added table entry and right contains the values of the added entry.

If a table entry that already was present is altered during the re-reading or streaming read of a file, then the value of tpe will be Input::EVENT_CHANGED. In this case left contains the index of the changed table entry and right contains the values of the entry before the change. The reason for this is that the table already has been updated when the event is raised. The current value in the table can be ascertained by looking up the current table value. Hence it is possible to compare the new and the old values of the table.

If a table element is removed because it was no longer present during a re-read, then the value of tpe will be Input::EVENT_REMOVED. In this case left contains the index and right the values of the removed element.

Filtering data during import

The input framework also allows a user to filter the data during the import. To this end, predicate functions are used. A predicate function is called before a new element is added/changed/removed from a table. The predicate can either accept or veto the change by returning true for an accepted change and false for a rejected change. Furthermore, it can alter the data before it is written to the table.

The following example filter will reject adding entries to the table when they were generated over a month ago. It will accept all changes and all removals of values that are already present in the table.

Input::add_table([$source="denylist.file", $name="denylist",
                  $idx=Idx, $val=Val, $destination=denylist,
                  $mode=Input::REREAD,
                  $pred(tpe: Input::Event, left: Idx, right: Val) = {
                    if ( tpe != Input::EVENT_NEW ) {
                        return T;
                    }
                    return (current_time() - right$timestamp) < 30day;
                  }]);

To change elements while they are being imported, the predicate function can manipulate left and right. Note that predicate functions are called before the change is committed to the table. Hence, when a table element is changed (tpe is Input::EVENT_CHANGED), left and right contain the new values, but the destination (denylist in our example) still contains the old values. This allows predicate functions to examine the changes between the old and the new version before deciding if they should be allowed.

Broken input data

The input framework notifies you of problems during data ingestion in two ways. First, reporter messages, ending up in reporter.log, indicate the type of problem and the file in which the problem occurred:

#fields ts      level   message location
0.000000        Reporter::WARNING       denylist.file/Input::READER_ASCII: Did not find requested field ip in input data file denylist.file.   (empty)

Second, the Input::TableDescription and Input::EventDescription records feature an $error_ev member to trigger events indicating the same message and severity levels as shown above. The use of these events mirrors that of change events.

For both approaches, the framework suppresses repeated messages regarding the same file, so mistakes in large data files do not trigger a message flood.

Finally, the ASCII reader allows coarse control over the robustness in case of problems during data ingestion. Concretely, the InputAscii::fail_on_invalid_lines and InputAscii::fail_on_file_problem flags indicate whether problems should merely trigger warnings or lead to processing failure. Both default to warnings.

Reading Data to Events

The second data ingestion mode of the input framework directly generates Zeek events from ingested data instead of inserting them to a table. Event streams work very similarly to the table streams discussed above, and most of the features discussed (such as predicates for filtering) also work for event streams. To read the denylist of the previous example into an event stream, we use the Input::add_event function:

type Val: record {
    ip: addr;
    timestamp: time;
    reason: string;
};

event denylistentry(description: Input::EventDescription,
                     tpe: Input::Event, data: Val) {
    # do something here...
    print "data:", data;
}

event zeek_init() {
    Input::add_event([$source="denylist.file", $name="denylist",
                     $fields=Val, $ev=denylistentry]);
}

Event streams differ from table streams in two ways:

  • An event stream needs no separate index and value declarations — instead, all source data types are provided in a single record definition.

  • Since the framework perceives a continuous stream of events, it has no concept of a data baseline (e.g. a table) to compare the incoming data to. Therefore the change event type (an Input::Event instance, tpe in the above) is currently always Input::EVENT_NEW.

These aside, event streams work exactly the same as table streams and support most of the options that are also supported for table streams.

Data Readers

The input framework supports different kinds of readers for different kinds of source data files. At the moment, the framework defaults to ingesting ASCII files formatted in the Zeek log file format (tab-separated values with a #fields header line). Several other readers are included in Zeek, and Zeek packages/plugins can provide additional ones.

Reader selection proceeds as follows. The Input::default_reader variable defines the default reader: Input::READER_ASCII. When you call Input::add_table or Input::add_event this reader gets used automatically. You can override the default by assigning the $reader member in the description record passed into these calls. See test cases in testing/btest/scripts/base/frameworks/input/ for examples.

The ASCII Reader

The ASCII reader, enabled by default or by selecting Input::READER_ASCII, understands Zeek’s TSV log format. It actually understands the full set of directives in the preamble of those log files, e.g. to define the column separator. This is rarely used, and most commonly input files merely start with a tab-separated row that names the #fields in the input file, as shown earlier.

Warning

The ASCII reader has no notion of file locking, including UNIX’s advisory locking. For large files, this means the framework might process a file that’s still written to. The reader handles resulting errors robustly (e.g. via the reporter log, as described earlier), but nevertheless will encounter errors. In order to avoid these problems it’s best to produce a new input file on the side, and then atomically rename it to the filename monitored by the framework.

There’s currently no JSON ingestion mode for this reader.

The Benchmark Reader

The benchmark reader, selected via Input::READER_BENCHMARK, helps the Zeek developers optimize the speed of the input framework. It can generate arbitrary amounts of semi-random data in all Zeek data types supported by the input framework.

The Binary Reader

This reader, selected via Input::READER_BINARY, is intended for use with file analysis input streams to ingest file content (and is the default type of reader for those streams).

The Raw Reader

The raw reader, selected via Input::READER_RAW, reads a file that is split by a specified record separator (newline by default). The contents are returned line-by-line as strings; it can, for example, be used to read configuration files and the like and is probably only useful in the event mode and not for reading data to tables.

The SQLite Reader

The SQLite input reader, selected via Input::READER_SQLITE, provides a way to access SQLite databases from Zeek. SQLite is a simple, file-based, widely used SQL database system. Due to the transactional nature of SQLite, databases can be used by several applications simultaneously. Hence they can, for example, be used to make constantly evolving datasets available to Zeek on a continuous basis.

Reading Data from SQLite Databases

Like with Zeek’s logging support, reading data from SQLite databases is built into Zeek without any extra configuration needed. Just like text-based input readers, the SQLite reader can read data — in this case the result of SQL queries — into tables or events.

Reading Data into Tables

To read data from a SQLite database, we first have to provide Zeek with the information how the resulting data will be structured. For this example, we expect that we have a SQLite database, which contains host IP addresses and the user accounts that are allowed to log into a specific machine.

The SQLite commands to create the schema are as follows:

create table machines_to_users (
host text unique not null,
users text not null);

insert into machines_to_users values (
    '192.168.17.1', 'johanna,matthias,seth');
insert into machines_to_users values (
    '192.168.17.2', 'johanna');
insert into machines_to_users values (
    '192.168.17.3', 'seth,matthias');

After creating a file called hosts.sqlite with this content, we can read the resulting table into Zeek:

type Idx: record {
   host: addr;
};

type Val: record {
   users: set[string];
};

global hostslist: table[addr] of Val = table();

event zeek_init()
   {
   Input::add_table([$source="/var/db/hosts",
       $name="hosts",
       $idx=Idx,
       $val=Val,
       $destination=hostslist,
       $reader=Input::READER_SQLITE,
       $config=table(["query"] = "select * from machines_to_users;")
       ]);

   Input::remove("hosts");
   }

event Input::end_of_data(name: string, source: string)
   {
   if ( name != "hosts" )
       return;

   # now all data is in the table
   print "Hosts list has been successfully imported";

   # List the users of one host.
   print hostslist[192.168.17.1]$users;
   }

The hostslist table can now be used to check host logins against an available user list.

Turning Data into Events

The second mode is to use the SQLite reader to output the input data as events. Typically there are two reasons to do this. First, the structure of the input data is too complicated for a direct table import. In this case, the data can be read into an event which can then create the necessary data structures in Zeek in scriptland. Second, the dataset is too big to hold in memory. In this case, event-driven ingestion can perform checks on-demand.

As an example, let’s consider a large database with malware hashes. Live database queries allow us to cross-check sporadically occurring downloads against this evolving database. The SQLite commands to create the schema are as follows:

create table malware_hashes (
    hash text unique not null,
    description text not null);

insert into malware_hashes values ('86f7e437faa5a7fce15d1ddcb9eaeaea377667b8', 'malware a');
insert into malware_hashes values ('e9d71f5ee7c92d6dc9e92ffdad17b8bd49418f98', 'malware b');
insert into malware_hashes values ('84a516841ba77a5b4648de2cd0dfcb30ea46dbb4', 'malware c');
insert into malware_hashes values ('3c363836cf4e16666669a25da280a1865c2d2874', 'malware d');
insert into malware_hashes values ('58e6b3a414a1e090dfc6029add0f3555ccba127f', 'malware e');
insert into malware_hashes values ('4a0a19218e082a343a1b17e5333409af9d98f0f5', 'malware f');
insert into malware_hashes values ('54fd1711209fb1c0781092374132c66e79e2241b', 'malware g');
insert into malware_hashes values ('27d5482eebd075de44389774fce28c69f45c8a75', 'malware h');
insert into malware_hashes values ('73f45106968ff8dc51fba105fa91306af1ff6666', 'ftp-trace');

The following code uses the file-analysis framework to get the sha1 hashes of files that are transmitted over the network. For each hash, a SQL-query runs against SQLite. If the query returns a result, we output the matching hash.

@load frameworks/files/hash-all-files

type Val: record {
   hash: string;
   description: string;
};

event line(description: Input::EventDescription, tpe: Input::Event, r: Val)
   {
   print fmt("malware-hit with hash %s, description %s", r$hash, r$description);
   }

global malware_source = "/var/db/malware";

event file_hash(f: fa_file, kind: string, hash: string)
   {

   # check all sha1 hashes
   if ( kind=="sha1" )
       {
       Input::add_event(
           [
           $source=malware_source,
           $name=hash,
           $fields=Val,
           $ev=line,
           $want_record=T,
           $config=table(
               ["query"] = fmt("select * from malware_hashes where hash='%s';", hash)
               ),
           $reader=Input::READER_SQLITE
           ]);
       }
   }

event Input::end_of_data(name: string, source:string)
   {
   if ( source == malware_source )
       Input::remove(name);
   }

If you run this script against the trace in testing/btest/Traces/ftp/ipv4.trace, you will get one hit.