blogwright-analytics 0.4.0-beta.2 → 0.4.0-beta.4

package/dist/nodes.js

@@ -3,24 +3,29 @@
  * CloudFront-logs-to-Iceberg pipeline is built from and nothing else: the
  * site's own bucket, distribution and log group stay in the CLI's graph
  * (`packages/cli/src/nodes.ts`) and are never touched from here. This module
- * carries all twelve of them, in four chains. The table chain - the S3 Tables bucket,
+ * carries all fourteen of them, in four chains. The table chain - the S3 Tables bucket,
  * the namespace inside it, the `page_views` table, and the Glue federation
  * Firehose reads that table through - runs `analytics-table-bucket` ->
  * `analytics-namespace` -> `analytics-table` ->
  * `analytics-catalog-integration`. The transform chain - the long-lived
- * `visitor_key` salt, the Lambda execution role whose policy names that
- * secret's ARN, and the record-transform function itself - runs
- * `analytics-salt-secret` -> `analytics-transform-role` ->
- * `analytics-transform-function`. The delivery chain - the bucket every record
- * Firehose cannot deliver lands in, the role it assumes, and the stream itself
- * - runs `analytics-error-bucket` -> `analytics-firehose-role` ->
- * `analytics-firehose-stream`, and joins the other two chains through the
- * role's four grants and the stream's destination. The vended-delivery chain -
+ * `visitor_key` salt, the log group the function's own output lands in, the
+ * Lambda execution role whose policy names that secret's ARN, and the
+ * record-transform function itself - runs `analytics-salt-secret` ->
+ * `analytics-transform-role` -> `analytics-transform-function`, with
+ * `analytics-transform-log-group` a second edge into that function. The
+ * delivery chain - the bucket every record Firehose cannot deliver lands in,
+ * the log group Firehose's delivery errors land in, the role it assumes, and
+ * the stream itself - runs `analytics-error-bucket` ->
+ * `analytics-firehose-role` -> `analytics-firehose-stream`, with
+ * `analytics-firehose-log-group` a second edge into that stream, and joins the
+ * other two chains through the role's five grants and the stream's
+ * destination. The vended-delivery chain -
  * the CloudWatch delivery destination pointing at that stream and the delivery
  * joining it to the site's log source - runs `analytics-log-destination` ->
  * `analytics-log-delivery` and hangs off the stream. All four are wired through
  * `dependsOn`, and a node depends on every node whose recorded ARN it
- * interpolates. {@link buildAnalyticsNodes} at the foot of this module returns
+ * interpolates - plus the two writers whose log group has to exist before the
+ * first line they write, which is an ordering with nothing read back. {@link buildAnalyticsNodes} at the foot of this module returns
  * the assembled set, and `plugin.ts` hands it to the SPI's `Plugin.nodes`;
  * assembling an array is all it does - nothing here reconciles anything.
  *
@@ -41,8 +46,8 @@
  * over the host's `signingUsEast1` signer; no node here picks a region for a
  * request. {@link ANALYTICS_REGION} below is the same region as *text*, needed
  * only because an ARN spells its region out and because every node `title`
- * states the pin, so the bootstrap output an operator reads carries it. Ten of
- * the twelve titles state it as the region they are created in; the two IAM
+ * states the pin, so the bootstrap output an operator reads carries it. Twelve
+ * of the fourteen titles state it as the region they are created in; the two IAM
  * role nodes state it as the pipeline they serve, because IAM is global and
  * "created in us-east-1" is not a property a role has (§Region pinning says so
  * in as many words) - a title claiming otherwise would be the pin stated
@@ -85,16 +90,30 @@ const TABLE_NODE = 'analytics-table';
 const CATALOG_NODE = 'analytics-catalog-integration';
 /** The `analytics-salt-secret` node id, shared by its `id`, its state key and the edge into it. */
 const SALT_SECRET_NODE = 'analytics-salt-secret';
+/**
+ * The `analytics-transform-log-group` node id. The group the transform Lambda
+ * writes its own output to, owned here rather than left to the implicit
+ * creation Lambda would otherwise do on first invocation - which is what makes
+ * its retention a property at all.
+ */
+const TRANSFORM_LOG_GROUP_NODE = 'analytics-transform-log-group';
 /** The `analytics-transform-role` node id. */
 const TRANSFORM_ROLE_NODE = 'analytics-transform-role';
 /** The `analytics-transform-function` node id. */
 const TRANSFORM_FUNCTION_NODE = 'analytics-transform-function';
 /** The `analytics-error-bucket` node id, shared by its `id`, its state key and the edge into it. */
 const ERROR_BUCKET_NODE = 'analytics-error-bucket';
+/**
+ * The `analytics-firehose-log-group` node id. The group Firehose writes its
+ * delivery errors to, and the one node in this module that owns a log stream as
+ * well as a group: enabling error logging through the API requires both to
+ * exist in advance.
+ */
+const FIREHOSE_LOG_GROUP_NODE = 'analytics-firehose-log-group';
 /** The `analytics-firehose-role` node id. */
 const FIREHOSE_ROLE_NODE = 'analytics-firehose-role';
 /**
- * The `analytics-firehose-stream` node id. Exported, alone among the twelve,
+ * The `analytics-firehose-stream` node id. Exported, alone among the fourteen,
  * because `analytics status` reads this node's recorded outputs back out of
  * the scoped state its `read` hydrated - the stream's delivery health - and a
  * second copy of the string in `commands.ts` would be a state key with two
@@ -649,9 +668,12 @@ const ZIP_MTIME = REPRODUCIBLE_ZIP_MTIME;
 /** The deflate level `packageAndUploadAgent` uses. */
 const ZIP_LEVEL = 6;
 /**
- * The prefix Lambda derives a function's log group from. The group itself is
- * created by the Lambda service on first invocation and by no node in this
- * graph - see {@link transformLogGroupArn}.
+ * The prefix Lambda derives a function's log group from, and therefore the
+ * prefix {@link transformLogGroupName} builds the one group this plugin owns
+ * for its function from. `analytics-transform-log-group` creates that group; it
+ * is not left to the implicit creation the Lambda service performs on a
+ * function's first invocation, because a group created that way is retained
+ * forever and carries none of the environment's tags.
  */
 const LAMBDA_LOG_GROUP_PREFIX = '/aws/lambda/';
 /** The name of the inline policy this plugin puts on its own transform role. */
@@ -763,10 +785,25 @@ function transformFunctionName(ctx) {
 function transformRoleName(ctx) {
     return boundedName(`${ctx.names.prefix}${TRANSFORM_ROLE_SUFFIX}`, ROLE_NAME_MAX_LENGTH, 'transform role');
 }
+/**
+ * The transform Lambda's log group name - **the one home this string has.**
+ *
+ * `analytics-transform-log-group` creates the group under this name and
+ * {@link transformLogGroupArn} scopes the role's grant to the ARN built from
+ * it, so the group that exists and the group the function is allowed to write
+ * to are one string by construction rather than two literals that happen to
+ * agree. Derived from {@link transformFunctionName} rather than re-derived from
+ * `ctx.names.prefix` for the same reason: Lambda's own group name is that
+ * function's name under {@link LAMBDA_LOG_GROUP_PREFIX}, so a second derivation
+ * could name a group Lambda never writes to.
+ */
+function transformLogGroupName(ctx) {
+    return `${LAMBDA_LOG_GROUP_PREFIX}${transformFunctionName(ctx)}`;
+}
 /**
  * The log group ARN the role's `logs:` grant is scoped to - the function's
- * **own** group and nothing else, the scoping
- * `packages/cli/src/nodes.ts:212` applies to the site's exec role.
+ * **own** group and nothing else, the scoping the site's exec role applies
+ * (`packages/cli/src/nodes.ts`).
  *
  * The region is {@link ANALYTICS_REGION} and not `ctx.config.region`, which is
  * the one place this ARN differs from the CLI's `logGroupArn` helper (whose
@@ -775,18 +812,31 @@ function transformRoleName(ctx) {
  * log group is too, and a grant naming the primary region would be a grant on a
  * group that never exists.
  *
- * **No node creates this group.** Lambda creates it implicitly on the
- * function's first invocation. That is worth stating because the policy below
- * grants `logs:CreateLogStream` and `logs:PutLogEvents` and *not*
- * `logs:CreateLogGroup`: the transform's own diagnostics therefore depend on
- * that implicit creation succeeding, and the pipeline's real failure signal is
- * elsewhere - a record the transform cannot map goes to Firehose's error prefix
- * (`transform/handler.ts`), and a batch that throws raises Firehose's own error
- * metric. Adding the group as a node of its own, with the retention the site's
- * log groups carry, is a coherent follow-up and is outside this node set.
+ * **`analytics-transform-log-group` creates this group**, which is why the
+ * policy below grants `logs:CreateLogStream` and `logs:PutLogEvents` and *not*
+ * `logs:CreateLogGroup`: the role has nothing to create, the shape the site's
+ * exec role has. An earlier version of this comment said no node creates the
+ * group and that Lambda's implicit creation on the function's first invocation
+ * was enough; production disproved it - the group never appeared, the transform
+ * ran and reported nothing, and the pipeline's record-level signals (a record
+ * the transform cannot map goes to Firehose's error prefix,
+ * `transform/handler.ts`, and a batch that throws raises Firehose's own error
+ * metric) answered *which* and never *why*.
  */
 function transformLogGroupArn(ctx) {
-    const group = `${LAMBDA_LOG_GROUP_PREFIX}${transformFunctionName(ctx)}`;
+    return analyticsLogGroupArn(ctx, transformLogGroupName(ctx));
+}
+/**
+ * A log group's ARN in {@link ANALYTICS_REGION}, in the `:*` form both an IAM
+ * grant and this plugin's recorded outputs use.
+ *
+ * Takes no region parameter, which is the one way it differs from the CLI's own
+ * `logGroupArn` (`packages/cli/src/nodes.ts`) - that helper defaults to
+ * `ctx.config.region`, correctly, because the site's groups live there. Every
+ * group this plugin names lives in us-east-1 with the rest of the pipeline, so
+ * there is no region for a caller to choose and no default for one to forget.
+ */
+function analyticsLogGroupArn(ctx, group) {
     return `arn:aws:logs:${ANALYTICS_REGION}:${ctx.accountId}:log-group:${group}:*`;
 }
 /**
@@ -856,8 +906,13 @@ function requireTransformRoleArn(ctx) {
  * without a teardown.
  *
  * **Two statements, two concrete resources, no `*` anywhere.** The `logs`
- * statement names the function's own log group ({@link transformLogGroupArn});
- * the `secretsmanager` statement names the one secret this pipeline owns and
+ * statement names the function's own log group ({@link transformLogGroupArn})
+ * and carries two actions rather than three: `logs:CreateLogGroup` is
+ * deliberately absent, because `analytics-transform-log-group` creates that
+ * group and a role granted the creation of a group it never has to create is a
+ * grant with no call behind it - the same shape the site's exec role has. What
+ * Lambda needs at runtime is the stream inside that group and the events in it.
+ * The `secretsmanager` statement names the one secret this pipeline owns and
  * nothing else. A `*` in the second would hand every secret in the account -
  * every other environment's salt, and `blogwright-pds`'s OAuth client key and
  * live session - to a role whose only job is to read one value, and nothing in
@@ -1024,6 +1079,100 @@ export function analyticsSaltSecretNode() {
         },
     };
 }
+/**
+ * A CloudWatch log group this plugin owns, on the read/create/update/delete
+ * contract the site's own groups have (`logGroupNode`,
+ * `packages/cli/src/nodes.ts`): `read` reports presence and records the ARN,
+ * `create` ensures the group with the environment's tags and then applies
+ * {@link LOG_RETENTION_DAYS}, `update` re-applies that retention on every
+ * apply, and `delete` removes the group.
+ *
+ * Owning the group is what makes retention a property at all. A group a service
+ * creates for itself is retained **forever** and no reconcile ever notices;
+ * re-applying the policy on every `update` also converts a group that already
+ * exists in that state - an environment provisioned before these two nodes
+ * existed - without a teardown.
+ *
+ * `stream`, when given, is a log stream the group must also hold. It is ensured
+ * on `create` **and re-ensured on every `update`**, which is the one place these
+ * nodes depart from the site's `logGroupNode` and is the
+ * {@link applyErrorBucketConfiguration} reconcile-on-every-apply pattern rather
+ * than the site's narrower update. The reason is concrete: `read()` answers on
+ * the group alone, so a group left by a run that stopped between
+ * `CreateLogGroup` and `CreateLogStream` reports present forever while an
+ * `update` that only re-applied retention would do nothing about the stream it
+ * is missing. Deleting a group deletes the streams inside it, so `delete` needs
+ * no counterpart.
+ *
+ * `dependsOn: []` on both: a group is the head of the chain that writes to it
+ * and reads no other node's output.
+ */
+function analyticsLogGroupNode(spec) {
+    const { id, title, name, stream } = spec;
+    const record = (ctx, group) => {
+        const out = output(ctx, id);
+        out.name = group;
+        out.arn = analyticsLogGroupArn(ctx, group);
+    };
+    return {
+        id,
+        dependsOn: [],
+        title,
+        async read(ctx) {
+            const group = name(ctx);
+            if (!(await logs(ctx).logGroupExists(group)))
+                return false;
+            record(ctx, group);
+            return true;
+        },
+        async create(ctx) {
+            const group = name(ctx);
+            const client = logs(ctx);
+            await client.ensureLogGroup(group, ctx.tags);
+            // Identity before the secondary calls, the ordering `analytics-error-bucket`
+            // and the site's own `bucketNode` follow: a crash between CreateLogGroup and
+            // the retention or the stream must still leave the group recorded in state.
+            record(ctx, group);
+            await client.putRetentionPolicy(group, LOG_RETENTION_DAYS);
+            if (stream !== undefined)
+                await client.ensureLogStream(group, stream);
+        },
+        async update(ctx) {
+            const group = name(ctx);
+            const client = logs(ctx);
+            await client.putRetentionPolicy(group, LOG_RETENTION_DAYS);
+            if (stream !== undefined)
+                await client.ensureLogStream(group, stream);
+        },
+        async delete(ctx) {
+            // `deleteLogGroup` swallows its own not-found, so a re-run after a
+            // completed teardown is a no-op rather than a failure.
+            await logs(ctx).deleteLogGroup(name(ctx));
+        },
+    };
+}
+/**
+ * The transform Lambda's own log group: the mapping decisions, the drop path,
+ * and the cold-start read of the salt secret - the only place this pipeline
+ * says *why* a record went where it did.
+ *
+ * Lambda writes into it under the execution role's existing
+ * `logs:CreateLogStream` and `logs:PutLogEvents`, scoped to this group and no
+ * other ({@link transformLogGroupArn}). The role is deliberately not granted
+ * `logs:CreateLogGroup`, because this node is what creates the group.
+ *
+ * It is at the head of the chain that writes to it -
+ * `analytics-transform-function` declares the edge - and the role does not,
+ * because the role's policy *derives* this group's ARN from the function's name
+ * rather than reading a recorded one, so there is no output to wait for.
+ */
+export function analyticsTransformLogGroupNode() {
+    return analyticsLogGroupNode({
+        id: TRANSFORM_LOG_GROUP_NODE,
+        title: `Transform Lambda log group (${ANALYTICS_REGION})`,
+        name: transformLogGroupName,
+    });
+}
 /**
  * The transform Lambda's execution role: permission to write its own logs and
  * to read the one secret it needs, and nothing else.
@@ -1310,12 +1459,16 @@ function whileRoleIsPropagating(call) {
  * artifact's name, even though the zip travels inline rather than through a
  * bucket (see {@link MAX_INLINE_ZIP_BYTES} for why inline).
  *
- * It depends on `analytics-transform-role`, whose recorded ARN it runs as.
+ * It depends on `analytics-transform-role`, whose recorded ARN it runs as, and
+ * on `analytics-transform-log-group`, whose ARN it reads nothing of: a group
+ * that does not exist when the function first runs is a log line lost with
+ * nothing raised, and on teardown the reverse order the engine walks removes
+ * the function before the group holding the evidence of what it did.
  */
 export function analyticsTransformFunctionNode() {
     return {
         id: TRANSFORM_FUNCTION_NODE,
-        dependsOn: [TRANSFORM_ROLE_NODE],
+        dependsOn: [TRANSFORM_ROLE_NODE, TRANSFORM_LOG_GROUP_NODE],
         title: `Record-transform Lambda (${ANALYTICS_REGION})`,
         async read(ctx) {
             const name = transformFunctionName(ctx);
@@ -1474,6 +1627,41 @@ const FIREHOSE_TRUST = {
  * that has no second case on this side of it.
  */
 const ERROR_OUTPUT_PREFIX = 'firehose-errors/';
+/**
+ * Days both of this plugin's log groups retain what is written to them,
+ * re-applied on every `update` the way the site's own groups re-apply theirs.
+ *
+ * 365, matching the site's `retention.microvmDays` default - a year of the
+ * build's own output, and a year of this pipeline's. It is a plugin-owned
+ * constant and **not** read from `ctx.config.retention`, deliberately: that
+ * block holds exactly two keys, `microvmDays` and `cloudfrontDays`, and each
+ * one names one of the site's own two log groups. A third and fourth consumer
+ * reading either would make an operator's setting for a group they named
+ * silently govern two resources it was never named for, so that an environment
+ * shortening its CloudFront retention would shorten the transform's diagnostics
+ * with it and nothing would say so. Making retention configurable per plugin is
+ * a config change with a name of its own, not a key borrowed here.
+ */
+const LOG_RETENTION_DAYS = 365;
+/**
+ * The log stream inside {@link firehoseLogGroupName} that Firehose writes its
+ * delivery errors to. Firehose's own name for that stream - enabling error
+ * logging through the API names it explicitly, and the service creates neither
+ * the group nor the stream.
+ *
+ * `BackupDelivery`, the stream Firehose uses for a destination configured with
+ * S3 backup, is deliberately not created: the Iceberg destination this pipeline
+ * builds configures none, so a second stream would be an empty one forever.
+ */
+const DESTINATION_DELIVERY_STREAM = 'DestinationDelivery';
+/**
+ * The prefix a Firehose delivery stream's log group is conventionally named
+ * under, the counterpart of {@link LAMBDA_LOG_GROUP_PREFIX}. Firehose derives
+ * nothing from it - it writes to whatever group the stream's
+ * `CloudWatchLoggingOptions` names - so the convention is this plugin's to keep,
+ * and {@link firehoseLogGroupName} is where it is kept.
+ */
+const FIREHOSE_LOG_GROUP_PREFIX = '/aws/kinesisfirehose/';
 /**
  * Seconds Firehose buffers records before writing a file, and the size in MiB
  * that would flush one sooner. Both are sent, because the service requires the
@@ -1525,6 +1713,39 @@ function firehoseRoleName(ctx) {
 function streamName(ctx) {
     return boundedName(`${ctx.names.prefix}${FIREHOSE_STREAM_SUFFIX}`, STREAM_NAME_MAX_LENGTH, 'delivery stream');
 }
+/**
+ * The log group Firehose writes its delivery errors to - **the one home this
+ * string has**, for {@link transformLogGroupName}'s reason and with more riding
+ * on it. `analytics-firehose-log-group` creates the group under this name
+ * today; the delivery role's grant on the stream inside it and the stream's own
+ * `CloudWatchLoggingOptions` are the two readers that follow, and each is meant
+ * to reach the name through this helper rather than spell a third literal of
+ * it. Firehose writes to whatever group its logging options name, so three
+ * spellings would fail as an empty group rather than as an error.
+ *
+ * Derived from {@link streamName} rather than from `ctx.names.prefix`, so the
+ * group is named after the stream whose errors it holds even if that stream's
+ * suffix ever changes.
+ */
+function firehoseLogGroupName(ctx) {
+    return `${FIREHOSE_LOG_GROUP_PREFIX}${streamName(ctx)}`;
+}
+/**
+ * The ARN of the one log stream Firehose writes its delivery errors to, which
+ * the delivery role's `logs:PutLogEvents` grant is scoped to.
+ *
+ * Not {@link analyticsLogGroupArn}'s `:*` form, and the difference is the whole
+ * point of a separate helper: the group ARN's trailing wildcard grants every
+ * stream the group will ever hold, while `PutLogEvents` authorises against a
+ * stream ARN and this role writes to exactly one. Both halves come from the
+ * same two helpers the log-group node and the stream's own
+ * `CloudWatchLoggingOptions` reach for, so the grant cannot name a stream that
+ * is not the one being written to.
+ */
+function firehoseLogStreamArn(ctx) {
+    const group = firehoseLogGroupName(ctx);
+    return `arn:aws:logs:${ANALYTICS_REGION}:${ctx.accountId}:log-group:${group}:log-stream:${DESTINATION_DELIVERY_STREAM}`;
+}
 /**
  * The error bucket's ARN as `analytics-error-bucket` recorded it. See
  * {@link requireRecordedArn}.
@@ -1637,13 +1858,13 @@ function glueGrantResources(ctx) {
  * and a table recreated under a new generated ARN reaches the policy without a
  * teardown.
  *
- * **Exactly four statements, one per capability the change spec names, every
+ * **Exactly five statements, one per capability the change spec names, every
  * `Resource` a concrete ARN and none of them `*`.** The action lists are AWS's
  * own, from the "Grant Firehose access to Amazon S3 Tables" policy under IAM
  * access control; what is narrowed is the resources, which that policy writes
  * with wildcards over the whole account.
  *
- * Two of the four are easy to get subtly wrong and are worth stating:
+ * Three of the five are easy to get subtly wrong and are worth stating:
  *
  * - the error-bucket statement names the bucket **and** `<bucket>/*`. Bucket
  *   actions (`s3:ListBucket`, `s3:GetBucketLocation`) authorise against the
@@ -1658,11 +1879,19 @@ function glueGrantResources(ctx) {
  *   unqualified invoke, so copying it would deny every transform call and send
  *   every record to the error bucket.
  *
- * There is no fifth statement. AWS's policy carries three more - Kinesis (this
- * stream is `DirectPut`), KMS (no customer-managed key is configured anywhere
- * in this pipeline) and CloudWatch Logs (no `CloudWatchLoggingOptions` is sent,
- * so Firehose writes no log stream to grant on) - and each of the three is
- * conditional on a feature this pipeline does not use.
+ * - the CloudWatch Logs statement names the **one log stream** Firehose writes
+ *   its delivery errors to, `<firehose log group>:log-stream:DestinationDelivery`,
+ *   and not the group's `:*` form. `logs:PutLogEvents` authorises against the
+ *   stream ARN, and this role writes to exactly one stream, so the wildcard the
+ *   group ARN carries would grant every stream a future group ever holds. Only
+ *   `PutLogEvents`: `analytics-firehose-log-group` creates the group *and* the
+ *   stream, so the role has nothing to create - the shape
+ *   `analytics-transform-role`'s grant has, one action shorter.
+ *
+ * That fifth statement is the only one of the three AWS's own policy adds that
+ * this pipeline needs. The other two stay out: Kinesis (this stream is
+ * `DirectPut`) and KMS (no customer-managed key is configured anywhere in this
+ * pipeline), each conditional on a feature this pipeline does not use.
  */
 async function applyFirehoseRolePolicy(ctx) {
     const errorBucketArn = requireErrorBucketArn(ctx, FIREHOSE_ROLE_NODE);
@@ -1720,6 +1949,14 @@ async function applyFirehoseRolePolicy(ctx) {
                 ],
                 Resource: [errorBucketArn, `${errorBucketArn}/*`],
             },
+            {
+                // Without this the destination's `CloudWatchLoggingOptions` are inert:
+                // Firehose is told where to write its delivery errors and is not allowed
+                // to, so the failure it was meant to explain stays as silent as before.
+                Effect: 'Allow',
+                Action: ['logs:PutLogEvents'],
+                Resource: firehoseLogStreamArn(ctx),
+            },
         ],
     });
 }
@@ -1754,6 +1991,12 @@ function firehoseDestination(ctx) {
         bufferIntervalSeconds: STREAM_BUFFER_INTERVAL_SECONDS,
         bufferSizeMb: STREAM_BUFFER_SIZE_MB,
         transformLambdaArn: requireTransformFunctionArn(ctx, FIREHOSE_STREAM_NODE),
+        // Both through the helpers `analytics-firehose-log-group` creates the group and
+        // the stream under, and the delivery role grants on - one string each. Firehose
+        // writes to whatever group its logging options name and creates nothing, so a
+        // second spelling of either would fail as an empty group rather than as an error.
+        logGroupName: firehoseLogGroupName(ctx),
+        logStreamName: DESTINATION_DELIVERY_STREAM,
     };
 }
 /**
@@ -1775,14 +2018,39 @@ function recordOptional(out, key, value) {
     else
         out[key] = value;
 }
+/**
+ * What the reconcile is about to change, named for the operator reading the log
+ * line, and only what actually differs.
+ *
+ * A single "AppendOnly <recorded> -> <desired>" would report a transition on
+ * every reconcile, including the logging-only one where `AppendOnly` is `true`
+ * on both sides - "AppendOnly true -> true", an operator told the reason for a
+ * call is a field that did not move. The caller reaches this only past a guard
+ * that returned on both flags matching, so at least one clause always fires and
+ * the result is never empty.
+ */
+function destinationDrift(appendOnly, loggingEnabled) {
+    const parts = [];
+    if (appendOnly !== STREAM_APPEND_ONLY) {
+        parts.push(`AppendOnly ${String(appendOnly)} -> ${String(STREAM_APPEND_ONLY)}`);
+    }
+    if (loggingEnabled !== true) {
+        parts.push(`error logging ${loggingEnabled === false ? 'off' : 'unrecorded'} -> on`);
+    }
+    return parts.join(', ');
+}
 /**
  * Record the delivery stream's identity and health from a `DescribeDeliveryStream`.
  *
  * `state` and `failure` are what `analytics status` reports (task 55), so the
  * stream's health is hydrated by the same `read` the reconcile runs and there is
  * no second describe path. `versionId` and `destinationId` are what
- * `UpdateDestination` cannot be called without, and `appendOnly` is the live
- * flag the reconcile compares against {@link STREAM_APPEND_ONLY}.
+ * `UpdateDestination` cannot be called without, and `appendOnly` and
+ * `loggingEnabled` are the two live flags the reconcile compares - the first
+ * against {@link STREAM_APPEND_ONLY}, the second against error logging simply
+ * being on. Both go through {@link recordOptional}, so a describe that stops
+ * reporting one clears it rather than leaving a stale `true` that would make the
+ * reconcile skip work it should do.
  *
  * The ARN is guarded on its value, the guard `analytics-table` and
  * `analytics-catalog-integration` both put on theirs: `describeDeliveryStream`
@@ -1797,6 +2065,7 @@ function recordStream(ctx, status) {
     recordOptional(out, 'versionId', status.versionId);
     recordOptional(out, 'destinationId', status.destinationId);
     recordOptional(out, 'appendOnly', status.appendOnly);
+    recordOptional(out, 'loggingEnabled', status.loggingEnabled);
     recordOptional(out, 'failure', status.failure);
 }
 /**
@@ -1945,10 +2214,39 @@ async function applyErrorBucketConfiguration(ctx, name) {
     await client.putBucketTagging(name, ctx.tags ?? {});
     await client.putPublicAccessBlock(name);
 }
+/**
+ * The log group Firehose writes its delivery errors to, and the log stream
+ * inside it that they are written to.
+ *
+ * **Firehose creates neither.** Enabling error logging through the API rather
+ * than the console requires the group *and* the stream to exist in advance, so
+ * this node creates both - and re-ensures the stream on every `update` beside
+ * the retention, which is why {@link analyticsLogGroupNode} takes a stream at
+ * all. Without that, a group left by a run that stopped between
+ * `CreateLogGroup` and `CreateLogStream` is permanently one call short, with
+ * `read()` reporting it present and `update()` doing nothing about it, and the
+ * symptom is a delivery failure with nowhere to be explained.
+ *
+ * {@link DESTINATION_DELIVERY_STREAM} is the only stream created:
+ * `BackupDelivery` belongs to a destination configured with S3 backup, and the
+ * Iceberg destination this pipeline builds configures none.
+ *
+ * It is at the head of the chain that writes to it -
+ * `analytics-firehose-stream` declares the edge - while the delivery role does
+ * not, for the reason `analytics-transform-log-group`'s counterpart does not.
+ */
+export function analyticsFirehoseLogGroupNode() {
+    return analyticsLogGroupNode({
+        id: FIREHOSE_LOG_GROUP_NODE,
+        title: `Firehose delivery-error log group (${ANALYTICS_REGION})`,
+        name: firehoseLogGroupName,
+        stream: DESTINATION_DELIVERY_STREAM,
+    });
+}
 /**
  * The role Firehose assumes to read the catalog, write the table, invoke the
- * transform and store what it could not deliver - four grants, four concrete
- * resources, no `*`.
+ * transform, store what it could not deliver and say so in one log stream -
+ * five grants, five concrete resources, no `*`.
  *
  * It declares `dependsOn` on the three nodes whose recorded ARNs those grants
  * interpolate. `topoSort` drains zero-indegree nodes alphabetically
@@ -2038,7 +2336,16 @@ export function analyticsFirehoseRoleNode() {
 export function analyticsFirehoseStreamNode() {
     return {
         id: FIREHOSE_STREAM_NODE,
-        dependsOn: [FIREHOSE_ROLE_NODE, TABLE_NODE, CATALOG_NODE, TRANSFORM_FUNCTION_NODE],
+        dependsOn: [
+            FIREHOSE_ROLE_NODE,
+            TABLE_NODE,
+            CATALOG_NODE,
+            TRANSFORM_FUNCTION_NODE,
+            // Not an ARN this node interpolates: the group has to exist before the
+            // stream can report a delivery failure into it, and on teardown the
+            // stream goes before the group that holds its errors.
+            FIREHOSE_LOG_GROUP_NODE,
+        ],
         title: `Firehose delivery stream (${ANALYTICS_REGION})`,
         async read(ctx) {
             const status = await firehose(ctx).describeDeliveryStream(streamName(ctx));
@@ -2049,10 +2356,11 @@ export function analyticsFirehoseStreamNode() {
             // "already exists", and the reconcile would go green over a stream that
             // accepts nothing.
             //
-            // What `update` then does with such a stream is *nothing*: it branches on
-            // the recorded `AppendOnly` flag alone, so a `CREATING_FAILED` or
-            // `DELETING` stream whose flag already matches is reconciled with zero AWS
-            // calls and reported done. That is stated rather than guarded because this
+            // What `update` then does with such a stream may be *nothing*: it branches
+            // on the two recorded flags and nothing else, so a `CREATING_FAILED` or
+            // `DELETING` stream that is already append-only with error logging on is
+            // reconciled with zero AWS calls and reported done - the state is not part
+            // of the comparison. That is stated rather than guarded because this
             // `read` is the hydration path - `recordStream` puts `state` and `failure`
             // into the plugin's scoped state, and reporting an unusable stream from
             // them is `analytics status`' job (task 55). Do not read this comment as a
@@ -2070,12 +2378,21 @@ export function analyticsFirehoseStreamNode() {
         async update(ctx) {
             const recorded = ctx.state.resources[FIREHOSE_STREAM_NODE];
             const appendOnly = typeof recorded?.['appendOnly'] === 'boolean' ? recorded['appendOnly'] : undefined;
-            // The live flag already matches what this pipeline wants, so there is
-            // nothing to reconcile and no AWS call at all. `undefined` does NOT match:
-            // a stream whose destination reported no flag, or a state file that lost
-            // it, is a stream this node cannot claim is append-only, and pushing the
-            // desired configuration is the safe direction.
-            if (appendOnly === STREAM_APPEND_ONLY)
+            const loggingEnabled = typeof recorded?.['loggingEnabled'] === 'boolean' ? recorded['loggingEnabled'] : undefined;
+            // Both live flags already match what this pipeline wants, so there is
+            // nothing to reconcile and no AWS call at all. `undefined` does NOT match
+            // either: a stream whose destination reported no flag, or a state file that
+            // lost one, is a stream this node cannot claim is configured, and pushing
+            // the desired configuration is the safe direction.
+            //
+            // **The logging half is what reaches the installed base.** Every stream
+            // this plugin has created is append-only, so on `appendOnly` alone this
+            // return fires for all of them and no already-deployed stream would ever
+            // be switched from silent delivery failures to logged ones. The second
+            // condition is what makes an existing stream reconcile exactly once: the
+            // update sends `CloudWatchLoggingOptions`, the re-read records
+            // `loggingEnabled: true`, and the next apply returns here.
+            if (appendOnly === STREAM_APPEND_ONLY && loggingEnabled === true)
                 return;
             const name = streamName(ctx);
             const client = firehose(ctx);
@@ -2098,7 +2415,7 @@ export function analyticsFirehoseStreamNode() {
                 // operator told the update was rejected when it succeeded.
                 let refusal;
                 try {
-                    ctx.logger.step(`updating the analytics delivery stream "${name}" in place (AppendOnly ${String(appendOnly)} -> ${STREAM_APPEND_ONLY}) - UpdateDestination keeps the stream's ARN, so the CloudFront log delivery pointed at it is untouched`);
+                    ctx.logger.step(`updating the analytics delivery stream "${name}" in place (${destinationDrift(appendOnly, loggingEnabled)}) - UpdateDestination keeps the stream's ARN, so the CloudFront log delivery pointed at it is untouched`);
                     await whileRoleIsPropagating(() => client.updateDestination(name, destination, { versionId, destinationId }));
                 }
                 catch (err) {
@@ -2637,6 +2954,7 @@ export function analyticsLogDeliveryNode() {
             // missing destination each fail with nothing sent.
             const site = requireSiteDeliverySource(ctx);
             const destinationArn = requireLogDestinationArn(ctx);
+            const requestDay = utcDay(new Date());
             await logs(ctx).createDelivery(site.source, destinationArn, {
                 // `CLOUDFRONT_RECORD_FIELDS`, never a list restated here: `schema.ts`
                 // owns which CloudFront fields exist, which ones fill a column, and
@@ -2669,7 +2987,7 @@ export function analyticsLogDeliveryNode() {
             // {@link CREATED_DAY_KEY} for why moving it later is the one direction
             // that corrupts data rather than merely losing some.
             if (typeof out[CREATED_DAY_KEY] !== 'string')
-                out[CREATED_DAY_KEY] = utcDay(new Date());
+                out[CREATED_DAY_KEY] = requestDay;
         },
         async delete(ctx) {
             // This plugin's own deliveries and nothing else - **never
@@ -2690,7 +3008,7 @@ export function analyticsLogDeliveryNode() {
     };
 }
 /**
- * The plugin's twelve resource nodes, assembled in the order the change spec's
+ * The plugin's fourteen resource nodes, assembled in the order the change spec's
  * §Analytics pipeline → Resource nodes table lists them. This is what
  * `Plugin.nodes` (`plugin.ts`) hands the CLI's generic `analytics bootstrap`
  * and `analytics destroy` verbs, and it is the whole of what this package
@@ -2711,7 +3029,7 @@ export function analyticsLogDeliveryNode() {
  *
  * **No `ctx` parameter, deliberately.** The SPI declares `nodes?(ctx)` and the
  * CLI calls it with one, so this function is assignable to it as written - a
- * zero-argument function satisfies a one-argument signature. None of the twelve
+ * zero-argument function satisfies a one-argument signature. None of the fourteen
  * factories needs a context to be *built*: each reads `ctx` inside `read`,
  * `create`, `update` and `delete`, when the reconcile is actually running. A
  * parameter accepted and ignored here would be an unused binding and, worse, a
@@ -2733,10 +3051,12 @@ export function buildAnalyticsNodes() {
         analyticsCatalogIntegrationNode(),
         // The transform chain.
         analyticsSaltSecretNode(),
+        analyticsTransformLogGroupNode(),
         analyticsTransformRoleNode(),
         analyticsTransformFunctionNode(),
         // The delivery chain.
         analyticsErrorBucketNode(),
+        analyticsFirehoseLogGroupNode(),
         analyticsFirehoseRoleNode(),
         analyticsFirehoseStreamNode(),
         // The vended-delivery chain.